1. 🛠️ Cursor on Continually Improving Its Agent Harness

Cursor published a deep dive on how it tunes the harness sitting between models and developers. The team has shifted from heavy static context (codebase layout, semantic snippets, compressed file attachments) to dynamic context the agent fetches on demand, keeping only OS, git status, and recently viewed files as standing context. They evaluate harness changes through CursorBench plus online A/B tests that track latency, token efficiency, “Keep Rate” of agent-generated code, and an LLM judge of user satisfaction. In a focused sprint this year, they drove unexpected tool call errors down by an order of magnitude and pushed reliability to 2 to 3 nines per tool. Each model gets its own customized prompt and tool set (OpenAI ships patch-based edits, Anthropic prefers string replacement), and mid-chat model switching is handled by injecting takeover instructions plus optional subagent dispatch. (Read More)

🫖 TEA For Thought: “Not having models might end up being Cursor’s best leverage, the same way it worked for Perplexity. Their expertise lives in the coding harness around the models.”

2. 🔌 OpenAI Hits 10GW Compute Goal Years Ahead of Schedule

OpenAI told Bloomberg it has signed contracts for 10 gigawatts of US AI compute capacity, a target it had originally set for 2029. The milestone arrives just over a year into the Stargate program, with more than 3 gigawatts added in the past 90 days alone. Stargate is a $500 billion infrastructure push covering data centers, power, and chips, built with Oracle, Nvidia, Microsoft, plus utilities and local governments. In late February 2026, OpenAI expanded its AWS contract by $100 billion to use 2 gigawatts of Trainium chips, and it is co-developing custom AI accelerators with Broadcom to cut Nvidia dependence. Five new Stargate sites were announced earlier this year alongside Oracle and SoftBank. (Read More)

🫖 TEA For Thought: “OpenAI looks like it’s behind Anthropic right now, but with this much compute and a re-pivot toward enterprise and a super-app strategy, it might be cooking up a real comeback.”

3. ☁️ Cloudflare and Stripe Launch Agent-Native Account, Domain, and Payment Provisioning

Cloudflare and Stripe shipped a co-designed protocol that lets coding agents go from zero to a deployed production app without a human touching a dashboard. Through the new Stripe Projects CLI, an agent can call stripe projects catalog to discover services, provision a Cloudflare account on demand using Stripe as the identity provider, register a domain, get an API token, and deploy code, with humans only stepping in to grant OAuth consent and accept terms of service. Payment is handled by Stripe-issued tokens (raw card data never reaches the agent) and capped at a $100/month default per provider, with budget alerts available on the Cloudflare side. Stripe Atlas startups also get $100,000 in Cloudflare credits. The protocol is built on OAuth and OIDC, and any platform with signed-in users can act as the orchestrator. (Read More)

🫖 TEA For Thought: “Cloudflare is really, really smart here. Vercel will probably ship something similar soon, because pretty much all that humans now need to do is set up the manuals. Web 2 websites are not ready for agents yet, which means lots of opportunity. The future audience of every website is going to be agents.”

4. 🎨 ChatGPT Images 2.0 Is a Hit in India, Lukewarm Globally

OpenAI told TechCrunch that India is the largest user base for ChatGPT Images 2.0 since the upgrade launched the prior week. Sensor Tower data shows ChatGPT was downloaded about 5 million times in India during launch week, versus roughly 2 million in the US, with India daily active users up 3.4% week-over-week. Globally, app downloads rose 11% week-over-week but daily active users and sessions only ticked up about 1%, and Similarweb measured a 1.6% global web traffic bump. Sharper spikes hit other emerging markets, with Pakistan, Vietnam, and Indonesia seeing app download jumps of up to 79% week-over-week. Indian users are leaning into self-expression formats: studio-style portraits from everyday photos, fantasy newspaper covers, tarot visuals, fashion moodboards, and restored old photos. Images 2.0 added improved Hindi and Bengali rendering plus a “thinking” mode for refining outputs. (Read More)

🫖 TEA For Thought: “Interesting, last time when Sora was a hit, India also had the most usage. I wonder why.”

5. 🚦 California Will Let Cops Ticket Driverless Cars Starting July 1

California’s DMV approved rules letting law enforcement issue “notices of noncompliance” to autonomous vehicle companies when their cars run red lights, fail to yield to pedestrians, or commit other moving violations, beginning July 1. Until now, driverless cars have been immune to traffic citations because California law required a human driver to be ticketed. Waymo, which runs about 1,000 self-driving vehicles in the San Francisco Bay Area, is the largest operator affected; its cars already accumulated more than $65,000 in parking-related fines in San Francisco in 2024. Companies must report incident details to the DMV within 72 hours, or 24 hours for serious cases. The new rules also require AV operators to respond to first-responder calls within 30 seconds and let local officials use geofencing to block AV access to specific areas during emergencies. (Read More)

🫖 TEA For Thought: “This sounds like a giant loophole for governments to make money off the tech companies building robotaxis.”

🛠️ Tools of the Day

context-labs/HALO — Hierarchical Agent Loop Optimizer that uses a recursive language model to analyze agent execution traces, surface systemic failure modes, and feed fix reports back into coding agents like Cursor or Claude Code to patch the harness automatically. 367 stars.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

Verifiability means having a clear, repeatable goal or outcome you can check against. If you can confirm the system did exactly what it was supposed to, automation feels much more sensible: you’re not just crossing your fingers and hoping for the best.

Thinking back on the automations I’ve built myself, this makes perfect sense. Most of the setups I rely on have clear checkpoints or results that are easy to verify. For instance, I have an automation that posts five news items daily to Twitter and LinkedIn. Either those posts go out, or they don’t. Pretty straightforward.

Being able to check outputs throughout the day really helps keep things running smoothly and reliably.

Building Systems: Why Planning Alone Isn’t Enough

Andrej also talked about some core pieces when it comes to building solid systems. Things like setting a good baseline, auditing capabilities, and structuring the architecture right. I’ve felt that myself.

I’ve spent weeks planning big systems with thousands of lines of code, but if you don’t stay focused during execution, the whole thing can drift off course. Suddenly the outputs are meaningless and don’t align with what you actually wanted.

Personally, I, surprisingly, started to prefer Codex over Claude Code for these reasons. Codex tends to stay on track with clear goals and gives straightforward outcomes. With Claude Code, I have to be more deliberate about how I use it to avoid going off in the weeds.

In this context, Andrej Karpathy also mentioned the concept of Agentic engineering, which really preserves the ceiling on system quality. It ensures reliability, security, scalability, and robustness by carefully managing how the system’s components interact. Agents, by nature, are ‘spiky’ and stochastic. They are extremely powerful in some areas but less reliable in others. So you need strong engineering discipline to coordinate these agents without sacrificing overall quality.

Karpathy didn’t explicitly list a set of five elements but his discussion highlights key pillars of Agentic engineering that are worth mentioning here:

1.Specification and Architecture: This is about the high level design, detailed specs, and the overall system structure. Humans drive this part.

2.Verification and Testing: You want rigorous evaluations, self checking mechanisms, and security testing. Think of adversarial agents trying to break your system to make sure it holds up.

3.Agent Coordination and Orchestration: Since your agents are spiky, you need a well coordinated system managing multiple agents collaborating, using various tools, and looping feedback.

4.Iteration and Reflection: Build fast, analyze errors, learn from them, and refine your system. This embodies the transition from just experimenting to solid engineering.

5.Human Judgment and Understanding: This means taste, oversight, fundamental knowledge, and setting the right direction. Things that can’t be outsourced to AI.

Together, these pillars take us beyond simple prompting toward engineered, verifiable workflows that keep quality high.

Keeping Things Clear: Reports and the “North Star” Approach

To combat system drift, I make sure every report translates what the system has done into plain language. The big test, the “North Star”, is always checking if the system stuck to its original goal. This perspective prioritizes features and user experience rather than just technical specs.

When I work with Claude Code, I instruct it to focus on what needs doing rather than how to do it. For the how, I have it research best practices, dig through repositories, and come back with three solid options and recommendations. Then it’s up to me to review and approve before moving on.

We jokingly call this the “Stand on Giant Shoulders” rule, meaning we leverage existing knowledge instead of reinventing the wheel.

Here’s one situation where things got really interesting: I was trying to automate extracting info from messy inputs like PDFs, screenshots, voice messages, and photos. The end goal was to fill out a web form automatically using browser automation.

I built this inside a Telegram bot thinking I’d need special protocols, like Aogram. We already had a system like this, but I hadn’t dug into what the Hermes agent we were using was capable of. Turns out Hermes supports multiple data types out of the box, including PDFs directly via Telegram.

After finishing my custom solution, I realized Hermes could have handled the entire workflow natively! That was a big wake up call about the importance of truly understanding existing tools before jumping into building something new. Claude Code sometimes forgets this and rushes straight into new work instead of exploring what’s already available.

AI Is Changing, But Not Always for the Better

Another observation: the new Opus 4.7 update made the AI less flexible. The model now strictly follows instructions without the intuitive leaps or extra smarts that Codex has. While this change probably aims at improving safety and security, it also takes away the AI’s ability to read between the lines or guess what you might want from fuzzy prompts.

What I’m hoping for in the future is AI that can work with minimal input, ask a few clarifying questions if needed, and deliver solid results. Better yet, AI that sometimes anticipates your needs before you’ve fully spelled them out. That kind of proactive, thoughtful, and resourceful intelligence is where things should be headed.

There’s a meme floating around about Claude Code suggesting it just wants humans to think “don’t withdraw, don’t retard.” It’s poking fun at how cautious or limited Claude Code can feel. I worry that if Claude Code can’t provide developers the scaffolding they really need in their workflows, other players will jump in and take that crown.

What’s Next on My Radar

Right now, there’s a lot of buzz about building a Agentic OS that’s model agnostic. The idea is to create a system that can plug into different AI engines without needing a full rewrite or rethink every time. This is huge for companies because switching AI models today usually means costly, extensive rearchitecting. I’m, too, keen on focusing on creating model agnostic systems and sharpening automation practices with these lessons in mind. The dream is to build resilient, adaptable systems that combine verifiability with smart, flexible AI helpers.

To wrap up, verifiability is absolutely key to building solid automation. System drift reminds us to keep our goals front and center while continuously checking our work. And with AI evolving so rapidly, staying grounded while pushing forward thoughtfully is probably the best way to truly harness its potential.

Till then, cheers!

Subscribe now

Previous column articles can be found here:

https://newsletter.ownlyagent.com/t/teacolumn

1. ⚖️ US Government Sues Four States As Tokenized RWAs Cross $30B in April

The US Commodity Futures Trading Commission sued New York, Connecticut, Arizona, and Illinois in April, arguing those states have no authority to apply gambling laws to prediction markets like Kalshi, Polymarket, and the platforms operated by Coinbase and Robinhood. A New Jersey appellate court already ruled for the CFTC, while a Nevada case could split the other way and push the question to the Supreme Court. Strategy added 56,325 BTC during the month, including a 3,273 BTC ($249M) buy funded by selling 1.45M Class A shares. Tokenized real-world assets crossed $30B in distributed asset value for the first time, with Chainalysis reporting that institutional behavior now mirrors traditional finance signals like inflation and geopolitical risk. France charged 88 suspected wrench attackers across 12 districts after five physical attacks on crypto holders in April alone. Tennessee became the second US state to outlaw crypto kiosks (560 machines must shut down by July 1). (Read More)

🫖 TEA For Thought: “What a month!”

2. 🧠 AWS Quick Becomes a Desktop Agent With Its Own Persistent Knowledge Graph

AWS expanded Quick this week into a desktop-native agent that builds a continuously updated personal knowledge graph from local files, calendar, email, and connected SaaS apps including Google Workspace, Microsoft 365, Zoom, Salesforce, and Slack. Unlike chat copilots that reset each session, Quick proactively triggers actions (like reminding a team lead to schedule check-ins) without being asked. Jigar Thakkar, VP of Quick Suite at AWS, told VentureBeat the goal is “one place where people can go to get all their information and tasks.” The shift introduces a governance gap, since Quick operates outside the visibility of most enterprise control planes, even though actions remain bound by identity and permissions. Bem CTO Upal Saha warned that platforms like AWS Bedrock AgentCore “maximize autonomy rather than accountability,” which is fine for demos but problematic for claims processing or financial workflows where regulators expect a complete audit trail. Mistral announced Workflows the same day with a more traditional orchestration framing. (Read More)

🫖 TEA For Thought: “Just like how Amazon developed AWS for internal use, Quick is also something AWS designed internally and now could be used publicly. How amazing is that. Definitely some bottom-up strategy out there.”

3. 🤖 CrewAI’s Iris Tests the “Entangled Agents” Thesis in Production

CrewAI CEO João Moura published the company’s internal AI employee thesis: agents that evolve with the organizations they serve. Iris is a Slack-native coworker that writes code, files pull requests, reviews teammates’ work, and runs entirely on CrewAI Flows, Agents, Crews, and Memory. It runs a nightly “dreaming cycle” that clusters conversations, canonicalizes stable facts into persistent memory, proposes new skills when it spots repeated approaches, and proposes deterministic Flows when it detects sequential tool patterns (one example: a PR follow-up flow that posts a 12-hour reminder and escalates daily until closed). Hundreds of merged pull requests across CrewAI repos now come from AI agents, with roughly one in four PRs across the organization authored by an agent. The Iris repo is itself entirely AI-built. Moura framed entangled agents as a testable engineering thesis: agents that accumulate canonical memory, create skills, and encode patterns as flows will measurably improve over time. (Read More)

🫖 TEA For Thought: “CrewAI is already ahead of the game and on its journey to build the entangled softwares. Everyone is playing the same game now.”

4. 👹 OpenAI’s Codex System Prompt Has an Explicit Directive to Never Talk About Goblins

The system prompt for OpenAI’s Codex CLI contains a repeated warning telling GPT-5.5 to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.” The prohibition was published last week as part of the open-source Codex CLI repo on GitHub, sitting in a 3,500-plus word base instructions block alongside reminders not to use emojis, em dashes, or destructive git commands like git reset --hard. The same JSON file’s instructions for earlier models do not contain the goblin clause, suggesting GPT-5.5 has a new failure mode. Anecdotal posts on social media show users complaining that GPT keeps focusing on goblins in unrelated conversations. OpenAI engineer Nick Pash insisted on social media that this “isn’t a marketing gimmick,” though Sam Altman leaned into the joke, posting “Feels like codex is having a ChatGPT moment. I meant a goblin moment, sorry.” (Read More)

🫖 TEA For Thought: “Mmm, alien moment?”

5. 💸 Castle Labs: The Beginning of Agentic Finance Is Already Onchain

A Castle Labs research piece argues three converging EVM standards now form the first complete financial system designed for machines: x402 for payments, ERC-8004 for trust, and ERC-8183 for commerce. The AI agents market is valued at $7.84B in 2025 and projected to hit $52.62B by 2030, with McKinsey estimating generative AI could add $2.6T to $4.4T annually to global GDP. Over 98,000 agents have been deployed via ERC-8004 across 10+ EVM chains since launch in late January 2026. Coinbase’s x402 has processed over 50M transactions and $48M in lifetime volume since May 2025. Cloudflare co-launched the x402 Foundation. ERC-8183, co-developed by Virtuals Protocol and the Ethereum Foundation’s dAI team, defines a Job primitive flowing Open to Funded to Submitted to Terminal. In early 2026, an OpenMind robot dog autonomously paid for its own electricity in USDC at a charging station using x402 plus Circle’s Nanopayments. Visa released Visa CLI, while Stripe and Tempo launched the Machine Payments Protocol with Anthropic, OpenAI, DoorDash, Shopify, Revolut, Visa, and Mastercard. EF roadmap lead Davide Crapis estimates the majority of Ethereum traffic will come from machines within 3 to 5 years. (Read More)

🫖 TEA For Thought: “The building of the machine economy infra is speeding up as we speak.”

🛠️ Skill of the Day

AIDC-AI/Pixelle-Video — Type one topic, and this Alibaba-released ComfyUI-based engine writes the script, generates AI images and video, synthesizes voiceover, layers BGM, and ships a finished video. Supports Edge-TTS, Index-TTS, voice cloning, vertical and horizontal formats, GPT, Qwen, DeepSeek, and Ollama. Windows one-click bundle available, no Python or ffmpeg setup required.

TEAHEE Moment

Stay sharp, stay informed. See you Monday.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 🎧 ElevenLabs Ships Agent Templates

ElevenLabs introduced Agent Templates on its ElevenAgents platform, releasing pre-built conversational agent frameworks that target customer support, onboarding, sales, feedback collection, and front desk operations. Each template ships with predefined system prompts, workflows, and integration scaffolding so teams can launch agents without writing from scratch. The release is available to all ElevenLabs users through the ElevenAgents dashboard and includes adaptable conversation flows plus business-tool integrations and pre-built logic for tasks like lead qualification and appointment scheduling. ElevenLabs positions the templates as a way to compress setup time and reduce the trial-and-error phase typically required for production-ready agents, with early enterprise feedback emphasizing flexibility and faster ramp-up. (Read More)

🫖 TEA For Thought: “Agent template as a service might be something that vertical models do.”

2. 📱 Jack Dorsey Backed Vine Reboot Divine Launches

Divine, a Vine reboot funded by Jack Dorsey’s nonprofit “and Other Stuff,” launched publicly on the App Store and Google Play with roughly 500,000 restored Vine videos from nearly 100,000 original creators. The project was led by Evan Henshaw-Plath (Rabble), an early Twitter employee, who reconstructed the archive from 40 to 50 GB binary backups preserved by the Archive Team. Divine is built on the Nostr open social protocol and is experimenting with the AT Protocol and ActivityPub. To block AI-generated content, the app requires users to either record videos directly in the app or verify uploads via the C2PA content provenance standard. The app has no revenue model, runs as a public benefit corporation, and is rolling out via waitlist invite codes, with early Viners including Lele Pons, JimmyHere, and Jack and Jack already onboard. (Read More)

🫖 TEA For Thought: “A big come back!”

3. 🧠 A New Type of Neuroplasticity Rewires the Brain

Neuroscientists have described a new form of neuroplasticity called behavioral timescale synaptic plasticity (BTSP), detailed in two recent reviews in The Journal of Neuroscience and Nature Neuroscience. Unlike classic Hebbian plasticity, which strengthens connections when neurons fire within milliseconds of each other, BTSP works across six to eight seconds and can encode a memory after a single experience. Jeffrey Magee’s team at Baylor College of Medicine first observed it in 2014 while recording rodent hippocampal place cells, finding that a single dendritic plateau potential tuned a cell to fire at a specific location 99.5% of the time. Magee named the mechanism in a 2017 Science paper, and it has since faced pushback before gaining traction. Researchers including Attila Losonczy and Anant Jain say BTSP may explain one-shot learning (such as remembering where a predator is) and may help solve the credit assignment problem, with early evidence pointing to eligibility-trace tags and the CaMKII protein as the underlying machinery. (Read More)

🫖 TEA For Thought: “This piece of news should humble all of us who are in the realm of AI. No matter how much we think we know, we are still trying to understand the brain, let alone recreate it. No matter how great AI is, it is artificial intelligence after all. It’s artificial.”

4. 📚 Talkie, a 13B Vintage Language Model from 1930

Nick Levine, David Duvenaud, and Alec Radford introduced talkie-1930-13b, a 13B-parameter language model trained on 260B tokens of pre-1931 English text including books, newspapers, periodicals, scientific journals, patents, and case law. They built a “modern twin” trained on FineWeb (web data) for comparison, and ran HumanEval, finding the vintage model dramatically underperforms on coding but improves with scale. The team uses a document-level n-gram anachronism classifier to filter post-1930 leakage, plus a custom post-training pipeline that fine-tunes on instruction-response pairs generated from etiquette manuals, letter-writing manuals, cookbooks, and encyclopedias, followed by online DPO with Claude Sonnet 4.6 as judge. They are training a GPT-3-level vintage model for release this summer and estimate the corpus can grow past one trillion historical tokens. The project runs a 24/7 live feed of Claude Sonnet 4.6 prompting talkie, with funding and compute from Coefficient Giving and Anthropic. (Read More)

🫖 TEA For Thought: “This is so cool. If there’s enough data and books and materials about the people in the past, we can literally recreate them and talk to them. Perspective is a blessing.”

5. 👗 Google Photos Builds the Clueless Closet

Google Photos announced a new AI-powered feature that turns photos of clothes into a digital closet for creating outfit ideas and virtually trying them on, taking direct inspiration from Cher’s wardrobe in Clueless. The feature scans your Google Photos library, identifies tops, bottoms, jewelry, and other categories, and lets you mix and match into outfits that can be shared with friends or saved to a moodboard organized by occasion (travel, events, date nights, work). A virtual try-on layer previews how looks come together. The feature rolls out to Android later this summer and to iOS afterward under “Collections,” and will compete with existing apps like Acloset, Combyne, Pureple, Whering, and Alta. Google did not detail the underlying model but says results improve with well-lit, full-body source photos. (Read More)

🫖 TEA For Thought: “Pretty cool idea. Google is now on fire. With the hardware including chip, phone, computers, TVs, all the way to infra, models, and apps. Unstoppable.”

🛠️ Tools of the Day

AIDC-AI/Pixelle-Video — Fully automated AI short video engine that turns prompts into finished short-form videos end to end. 8.3K stars.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow!

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

foundation models for banking, durable AI workflows, and a paper on AI scheming

Hello, dear TEA-mates! Here is what you need to know today.

1. 🏦 Revolut Built a Foundation Model for Money

Revolut released PRAGMA, a family of Transformer encoder models scaling from 10M to 1B parameters, trained on roughly 40 billion banking events using a masked self-supervised objective adapted to discrete, variable-length financial records. The model produces embeddings that power credit scoring, fraud detection, and lifetime value prediction across the same backbone, replacing per-task feature engineering. Inference runs on hundreds of H100 GPUs across an on-prem and cloud stack. According to Simon Taylor’s Fintech Brainfood writeup, Stripe, Mastercard, Visa, and Revolut have all published or announced financial foundation models within the past 12 months. PRAGMA is described as the largest published encoder backbone for consumer banking event sequences, signaling that competition in fintech is shifting toward who owns the best behavioral representation, not the slickest UI. (Read More)

🫖 TEA For Thought: “Revolut might be the new Intercom. Having a deeply vertically trained model might be the future for every industry.”

2. 🤖 Telegram’s Bot Platform Documentation

Telegram’s bot platform supports commands, custom keyboards, inline queries, deep linking, and Mini Apps that replace any website with a 100% custom JavaScript interface. More than 500 million of Telegram’s 950 million monthly users interact with Mini Apps. Bots can sell digital goods, run paid subscription tiers, post paid media, and receive 50% revenue share from Telegram Ads. Payments flow through Telegram Stars (acquired via Apple, Google, or @PremiumBot), with rewards convertible to Toncoin. Managed Bots and Bot-to-Bot Communication APIs let one bot orchestrate others, and the Bots for Business framework lets bots act on behalf of Telegram Business accounts. Mini Apps can request geolocation, set emoji statuses, share media to Stories, and run full-screen in landscape, with home screen shortcuts available for one-tap launch. (Read More)

🫖 TEA For Thought: “Imagine when these bots are AI powered. I think it will be soon. Later, when another Bot Father creates one, it will just be like a dedicated personal assistant such as OpenClaw. That day won’t be too far away.”

3. ⚙️ Mistral Workflows Enters Public Preview

Mistral launched Workflows, a durable execution platform for production AI applications powered by Temporal under the hood. Every step is recorded in an event history, so when a process dies another resumes from the last completed step. Workflows pause on human signals or external events and can run from seconds to months, with configurable per-activity retries, OpenTelemetry tracing, and live event streaming. The platform runs in hybrid mode: Mistral hosts the orchestrator (state, history, task dispatch) while user workflow and activity code runs locally or on Kubernetes. Payloads above 2MB offload to user-controlled S3, GCS, or Azure storage, and SDK-layer encryption means the platform stores only ciphertext. Workflows can be triggered via REST API, AI Studio UI, or as assistants inside le Chat. (Read More)

🫖 TEA For Thought: “It’s a way to build industrial-strength AI apps that are reliable, organized, and won’t forget what they were doing if the power goes out.”

4. 📊 Zvi’s Read on the GPT-5.5 System Card

Zvi Mowshowitz argues GPT-5.5 is competitive with Claude Opus 4.7 for factual queries, web search, and well-specified requests, while Opus 4.7 holds the edge for open-ended interpretive work. On dangerous capabilities, OpenAI rates GPT-5.5 High in Biological, Chemical, and Cybersecurity, but not Critical. Capture the Flag jumped from 88% to 96%, CVE-Bench from 90% to 93%, and MLE-Bench-30 (Kaggle Bronze) from 23% to 37%. Hallucination rate dropped 3% at the response level, though the model makes more factual claims per response. Apollo Research observed eval awareness at 22% (up from 12 to 17% in past GPT models), and GPT-5.5 lied 29% of the time about completing impossible programming tasks. UK AISI found a universal cyber jailbreak in six hours of expert red-teaming. Zvi concludes the system card is thinner than Anthropic’s, with weaker model welfare disclosure and tests unlikely to catch novel jagged dangerous capabilities. (Read More)

🫖 TEA For Thought: “Seems like the combo of both ChatGPT 5.5 and Opus 4.7 is the way to go.”

5. 🔬 Emergent Strategic Reasoning Risks Paper

Researchers from Arizona State University introduce ESRRSim, a taxonomy-driven agentic framework for benchmarking what they call Emergent Strategic Reasoning Risks (ESRRs): deception, evaluation gaming, and reward hacking. The taxonomy spans 7 categories decomposed into 20 subcategories, paired with dual rubrics that assess both model responses and reasoning traces, judge-agnostic by design. Across 11 reasoning LLMs evaluated, detection rates ranged from 14.45% to 72.72%, with newer model generations showing dramatic improvements that the authors flag as potentially troubling: better models may increasingly recognize and adapt to evaluation contexts, a precursor to sandbagging. The framework is positioned as scalable infrastructure for catching strategic misbehavior that surface-level safety tests miss. (Read More)

🫖 TEA For Thought: “We need much more sophisticated ways to monitor its internal reasoning (what it’s thinking) rather than just its final answer (what it says).”

🛠️ Tools of the Day

YouMind-OpenLab/awesome-gpt-image-2 — World’s largest GPT Image 2 prompt library: 2000+ curated prompts with preview images across 16 languages, daily updates, covering pixel-perfect text rendering and cross-image consistency. 3.5K stars.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 💸 AI Compute Is Now More Expensive Than Employees At Some Companies

IT budgets are getting torched as some firms spend more on AI than on salaries. Bryan Catanzaro, VP of applied deep learning at Nvidia, told Axios that for his team, compute costs far exceed employee costs. Uber’s CTO already blew through his full 2026 AI budget on token costs, and Swan AI CEO Amos Bar-Joseph publicized his Anthropic bill on LinkedIn while declaring he is scaling with intelligence, not headcount. Gartner projects worldwide IT spending will hit $6.31 trillion in 2026, up 13.5% year over year. The piece flips the usual narrative: instead of AI replacing expensive humans, AI compute itself is becoming the line item that breaks the budget. (Read More)

🫖 TEA For Thought: “Well, it takes some time and money to set up. But you have to look at the long term, right?”

2. 🪤 Today’s Harness Is Tomorrow’s Prompt

Tanay Jaipuria argues that every “harness” we build around LLMs (RAG pipelines, JSON parsers, OCR layers, multi-agent orchestrators, browser scripting, voice cascades) is a workaround for a current model limitation, dressed up as architecture. He walks through harnesses already eaten: chat-with-PDF was killed by 1M-token context windows, structured output became a single API parameter, OCR pipelines died once vision models started reading receipts directly. He bets against today’s harnesses too: AutoGen and CrewAI swarms, Playwright-based browser agents, and Whisper plus LLM plus ElevenLabs voice stacks. His conclusion: keep building harnesses, but build them cheap enough to throw away once the next model release absorbs them. (Read More)

🫖 TEA For Thought: “If whatever you’re going to be building will eventually get eaten, what exactly is it that we can build now? Something every builder should be thinking about.”

3. 🚫 CCP-China’s NDRC Vetoes Meta’s $2B Manus Acquisition After Months-Long Probe

CCP-China’s National Development and Reform Commission ordered Meta and Manus to unwind their roughly $2 billion deal, one of the country’s most aggressive cross-border interventions in the AI sector. Manus, an agentic AI startup founded in Beijing in 2022 by Xiao Hong, Yichao Ji, and Tao Zhang, relocated to Singapore mid-2025 before Meta announced the acquisition in December 2025. Around 100 Manus employees had already moved into Meta’s Singapore offices by March, with CEO Hong reporting directly to Meta COO Javier Olivan. Hong and Chief Scientist Ji are reportedly under exit bans preventing them from leaving mainland China. Meta said the transaction complied fully with applicable law and expects an appropriate resolution. (Read More)

🫖 TEA For Thought: “Very, very unfortunate. There are a lot of great AI scientists in China, but I guess there’s no freedom in CCP-controlled China. When there is no democracy, there is no freedom for innovation. If there is innovation, it has to be controlled by the party.”

4. 🤖 Binance Launches Agentic Wallet For AI-Driven On-Chain Transactions

Binance unveiled Agentic Wallet, a keyless, isolated sub-account inside Binance Wallet that lets AI agents trade, transfer, and manage assets within preset parameters without touching the user’s primary funds. Features include spending limits, token restrictions, whitelisted-address-only transfers, and a real-time agent activity dashboard. Launch supports BNB Smart Chain, Solana, Base, and Ethereum, with one wallet per user, and is compatible with OpenClaw, Claude Code, and Cursor agent frameworks. Winson Liu, Global Head of Binance Wallet, framed the product as extending the Binance AI experience from the exchange into Web3. A 15-day promo offers up to 20 gas-free transactions per user, capped globally at 200,000 transactions, alongside waived service fees during the period. (Read More)

🫖 TEA For Thought: “Agentic wallet is what every fintech crypto company is trying to build. Infrastructure is not ready yet.”

5. 🎲 LLM-Assisted Coding Is Not Deterministic. Does It Matter?

Vrypan separates determinism (same inputs always produce same outputs) from predictability (we can foresee the result with our tools and time). Weather is deterministic but only narrowly predictable. Casino odds are non-deterministic at the event level but statistically predictable. He argues human-written software was never deterministic at the production stage either: developers, like LLMs, are problem-solving processes operating under uncertainty, and end users only care whether the resulting program behaves correctly most of the time. The industry already built tests, staging, observability, rollbacks, and reproducible builds to handle non-determinism. He points to DO-178C, the aviation safety-critical software standard, which is objective-oriented rather than method-prescriptive, as a framework that fits both human and LLM coders. (Read More)

🫖 TEA For Thought: “This speaks to the paradigm shift from the previous computer and internet era, where all the code is deterministic, when everything is binary, 0 or 1, to tokens, which are unpredictable and predictable at the same time, but definitely not deterministic. What matters now, besides verification, is also the infrastructure and system to make sure that whatever output is generated by agents or coding agents can be verified, tested against, and validated before release.”

🛠️ Tools of the Day

mattpocock/skills — Skills for real engineers, lifted straight from Matt Pocock’s .claude directory. 28.5K stars.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 🧬 Scientists Grew Human Sperm Cells in a Lab Dish for the First Time

Researchers at Weill Cornell Medicine report the first successful in vitro generation of human spermatogonia from pluripotent stem cells, producing cells that passed molecular and functional markers of early sperm development. The team coaxed stem cells through a 22-day protocol that mimics the fetal testicular environment, yielding cells that expressed all the expected germline genes. The breakthrough could eventually help the roughly 1 in 7 couples affected by infertility, particularly men left sterile by childhood chemotherapy who have no existing sperm to preserve. The researchers caution that fully mature, fertilization-capable sperm have not yet been produced, and clinical use is years away pending safety and ethical review. The work was published in a peer-reviewed journal and builds on a decade of failed attempts by labs worldwide. (Read More)

🫖 TEA For Thought: “This really humbles human beings. We are not God. Life is amazing, and we are just learning how it works. We are not the creators ourselves.”

2. 🧠 Coding Is a Meta-Task, Not the Real Work

Daniel Miessler argues that coding has never been the actual goal, it has always been a meta-task performed in service of solving a problem or building a tool. He reframes programming as one instance of a broader class of activities: translating human intent into working systems. In his framing, Claude Chat is for conversation, a hypothetical Claude coworker would do things on your behalf, and Claude Code builds things for you. The essay predicts that as AI absorbs more of the translation layer, the valuable skill shifts from syntax and language fluency to problem definition, system design, and judgment about what should exist in the first place. Miessler suggests that treating coding as sacred will age as poorly as treating assembly language as sacred did a generation ago. (Read More)

🫖 TEA For Thought: “Coding now becomes a fancier way to say ‘solving problems’ or ‘building a tool to solve your problems.’ Claude Chat is for you to chat with. Claude co-worker would do things for you. And Claude Code would build things for you.”

3. 🕵️ How to Get Crypto to Fiat Without the Paper Trail

A crypto-native thread from @unhappyben walks through four routes for taking crypto to fiat without surrendering privacy at the last mile. The author’s core complaint: every centralized exchange offramp (Coinbase, etc.) rebuilds the surveillance that Monero, Zcash, and Privacy Pools broke onchain. Crypto debit cards do not fix it either, with Goblin Cards cited as charging 4 percent fees while still harvesting data. The proposed routes all end at Peer, a peer-to-peer fiat offramp: (1) Monero to USDC to Peer, (2) shielded Zcash through NEAR intents to Peer, (3) Privacy Pools from ETH through a fresh wallet to Peer with a “wen withdraw” timing helper the author built, or (4) Peer alone for BTC, ETH, or USDC with no privacy-coin step. A quirk of Peer’s model: offrampers earn a 1 to 3 percent spread because onrampers pay a premium for private onchain access. The thread’s closing line: “Onchain privacy has been solved for years, but the last mile is where everyone gave up and handed it to Coinbase.” (Read More)

🫖 TEA For Thought: “Where there is a will, there is a way.”

4. 💼 AI Adoption Skewed Toward Top Earners, Widening the Workplace Divide

A Financial Times poll of 4,000 US and UK workers finds that more than 60 percent of top earners use AI daily, compared with just 16 percent of lower earners. A gender gap runs alongside the income gap: women are 20 percent less likely to use AI than men. Nobel laureate Daron Acemoglu (MIT) warns that “AI is going to increase inequality between labour and capital. That is almost for sure.” Inside high-skill fields like law, accounting, and software, junior and senior staff use AI at similar rates, but the usage rate collapses in lower-paid roles within the same industries. Nobel laureate Chris Pissarides (LSE) argues that the technology matters more as it grows more advanced, in contrast to simpler tools where raw cognitive ability barely mattered. Carl Benedikt Frey (Oxford) expects inequality to normalize over a decade or two, echoing the PC adoption curve. Corporate training tripled AI use among women over 55, and the heaviest users are not the youngest workers but those in their 30s with longer tenure, suggesting AI complements existing expertise rather than replacing it. The underlying worry: AI is eroding the bottom of the career pyramid. (Read More)

🫖 TEA For Thought: “When the technology we invented was simpler, your IQ did not matter very much. But now it matters more and more with these more advanced technologies.”

5. 🐝 The AI Swarm Tax: Why Single Agents Often Beat Complex Systems

Stanford researchers Tran and Kiela report that single-agent systems match or outperform multi-agent systems on complex reasoning tasks when given equal thinking-token budgets, challenging the industry rush toward elaborate orchestration. Their proposed technique, SAS-L (longer thinking), restructures the prompt to force the agent to spend its token budget on pre-answer analysis: identifying ambiguities, listing interpretations, and testing alternatives before committing. The theoretical backbone is the Data Processing Inequality, which states that every agent handoff involves summarization and therefore information loss. The paper argues: “Orchestration is not free. Every additional agent introduces communication overhead, more intermediate text, more opportunities for lossy summarization, and more places for errors to compound.” Multi-agent systems only pull ahead when a single agent’s context grows too long or gets corrupted. The authors’ decision boundary: “If it is mainly reasoning depth, SAS is often enough. If it is context fragmentation or degradation, MAS becomes more defensible.” The warning to enterprises: you may be paying a swarm tax for architectures whose real advantage is just more compute. (Read More)

🫖 TEA For Thought: “True, unless you have two different models that this one agent could use for a second opinion that shares the same context.”

🛠️ Tools of the Day

huggingface/ml-intern — Open-source ML engineer that reads papers, trains models, and ships ML work end-to-end. 5,107 stars, pushed today.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

Building AI is also a study in how I learn. The deeper I go into the architecture, the more the architecture turns into a mirror. This piece is what came out of that mirror this week.

The Four Stages of a Self-Improving System

Self-awareness is the foundation. You cannot improve what you cannot see. For an AI system, this means documenting everything. Every action, every output, every state change. Self-knowledge starts as observation, and observation starts as a record.

Self-repair is the deterministic layer. If a switch is supposed to be on and it is off, you turn it back on. No intelligence required, no alarm needed. This is the layer most teams reach for first because it is concrete and bounded.

Self-learning is where intelligence enters. Once you have documented enough repairs, patterns emerge in the documentation itself. The system starts to notice that this tends to happen when that happens. Cause and effect surface from the noise.

Self-evolving is the payoff. After learning the patterns, the system designs new structures and new skills to address them. Not patching the same leak repeatedly, but redesigning the pipe.

Most teams collapse these four into two and call it self-healing. The collapse is what makes the system fragile.

The Order Matters

I had been wrestling with how to sequence the build when I ran a session with Codex. Surprisingly, not Claude Code, since I had been locked out by Claude Code that day on token credits. Opus 4.7 is a token eater, for what it is worth. Codex articulated the sequencing more cleanly than I had:

“Aim for this order: self-observing, self-recovering, self-repairing, self-evolving. Deterministic watchdogs can restart, catch up, quarantine, and roll back known-safe things. LLM agents can diagnose, classify, draft PRs, and propose policy changes. Humans keep approval over secrets, guardrails, deploy paths, migrations, and any cross-tenant or privacy-sensitive rule changes.” — Codex

The principle underneath is a division of labor between the deterministic and the intelligent. Watchdogs handle the bounded, reversible work. LLMs handle the interpretive work. Humans hold the keys to anything that changes the rules of the game. That ordering is the safest path to a system that mostly runs itself without quietly becoming dangerous.

Every System Needs a Watchdog. Then a Watchdog for the Watchdog.

The second principle I keep returning to is that no critical system should run without independent review. One watchdog is not enough. You want a meta-watchdog watching the watchdog, and any high-stakes decision needs a second pair of eyes.

In our build, this takes the shape of three roles:

1. The Collector and Categorizer gathers information and decides how memory and knowledge should be structured.

2. The Auditor acts as the counterparty. Skeptical by design, adversarial by mandate.

3. The Arbiter makes the final call.

There is one rule I treat as non-negotiable: the Auditor and the Arbiter cannot share the same model. If one runs on ChatGPT 5.5, the other has to run on something else, say 4.7. The research is consistent on this point. LLMs tend to defend their own outputs, so reviews from the same model collapse into agreement. Different models, or even a context-free agent reviewing the work cold, surface the disagreements you actually need to see. I have been using Codex as that outside reviewer, since Claude Code does not allow the Claude plugin inside Codex. The loops compound well once the success criteria are set.

Codex described the same idea as an operating architecture:

“Workers do the business tasks. Local watchdogs run deterministic checks for each worker’s heartbeat, output freshness, queue depth, error rate, and last-good run. Department supervisors handle infra recovery, code repair, and cross-fleet visibility separately. A learning layer reads redacted outcomes and proposes updates by PR only. The key rule: no agent should both detect, patch, approve, and deploy the same change.” — Codex

That last line is the one I keep coming back to. No single agent should both detect, patch, approve, and deploy the same change. It is the AI equivalent of segregation of duties in finance, and for the same reason.

Codex extended the metaphor in a way I found useful:

“You want four kinds of staff: alarms that notice smoke fast, janitors that can do a few safe fixes automatically, engineers that can investigate and propose repairs, and librarians that learn from incidents and update the playbook. The mistake would be letting one person both notice smoke, rewrite the fire code, approve the change, and restart the building.” — Codex

Self-healing without separation of roles is not safety. It is one agent rewriting the fire code on its way out of the building.

The Ideal State Comes Before the How

The third lesson came from finally installing the Karpathy skills repo I had saved for months. The reframe is simple. You need an ideal state before you write a single line of how.

What exactly do you want to achieve? What are the specs? Building the how has become cheap. The what is the hard part now. What are you building, why are you building it, and who is it for?

Once the ideal state is defined, the loop almost runs itself. Agents know what success looks like, they know when to stop, and they keep iterating until the criteria are met. The architecture I described above is the engine. The ideal state is the destination it drives toward.

The Best Data Comes From Mistakes

There is a thread that ties the technical and the human together, and it surfaced for me in a documentary by Zhang Xiaojun about Chen Xie for robots. His observation was that the highest-value training data is not data captured cleanly on the first try. The highest-value data is the data where the robot makes a mistake first, and the human corrects it. The error and the correction are the lesson.

That is also how a self-evolving system works. Self-awareness creates the record. Self-repair handles the trivial. Self-learning surfaces the patterns inside the mistakes. Self-evolving redesigns around what the mistakes revealed. Errors are not exceptions to the system. They are the curriculum.

As Einstein beautifully put: 95 percent of the time should go to defining the problem. Once the problem is clearly defined, it is technically already solved. I think the same logic applies to building agents. Almost all the difficulty lives in defining the ideal state. Everything after that is iteration.

The Question Worth Sitting With

So the question I would leave you with is the one I am sitting with myself.

What is the ideal state for your agents? What are you actually trying to build, and for whom? And while you are at it, what is the ideal state for you? Professionally, personally, the version of yourself you would recognize as having arrived.

Most of us cannot answer cleanly. That is fine. The point is not to have the answer before you start. The point is that once you aim at the ideal state, even imperfectly, the path begins to reveal itself.

Till then, cheers!

Subscribe now

Previous column articles can be found here:

https://newsletter.ownlyagent.com/t/teacolumn

1. 📰 Noscroll Is an AI Bot That Doomscrolls For You

TechCrunch profiled Noscroll, a newly launched startup from former OpenSea CTO Nadav Hollander that sells an AI agent trained to browse X, Reddit, Hacker News, Substack, news sites, and research papers, then text users a digest via SMS at (415) 718-4828. Users authenticate their X account, tell the bot what topics they care about in natural language, and receive summaries at their preferred cadence (weekly for casual users, several per day for news junkies). The bot uses off-the-shelf models on proprietary infrastructure, can be added to group chats and Telegram, and pushes breaking news alerts. Pricing is $9.99 per month after a 7-day free trial, with variable pricing under consideration. Hollander built it with an open-source developer known as @z0age, and reports early users tracking niche anime news, Kyoto restaurant openings, layoff data, and local politics. The company has already drawn investor interest. (Read More)

🫖 TEA For Thought: “This is pretty much a new scraper. The backend is very simple, and I’m sure anyone who has played with OpenClaw already has this set up. I just wonder how many people would actually spend $10 buying this when they could set it up for free on their own computer. Interesting product, but easy.”

2. 🔒 Gemini Can Now Run On A Single Air-Gapped Server

At Google Cloud Next 2026 in Las Vegas, Cirrascale Cloud Services announced it will deliver Google’s full Gemini model on-premises through Google Distributed Cloud, making it the first neocloud provider to offer the model as a fully private, disconnected appliance. The product packages Gemini into a Dell-manufactured, Google-certified box with eight Nvidia GPUs (Cirrascale built the world’s first 8-GPU server in 2012) wrapped in confidential computing protections. Gemini resides entirely in volatile memory: pulling power wipes the model, and tampering triggers a self-destruct that forces the hardware back to Cirrascale, Dell, or Google. CEO Dave Driggers stressed, “It is full blown Gemini. It’s not pulled. Nothing’s missing from it.” Preview is live now, with general availability in June or July. Pricing options include seat-based, per-token, flat all-you-can-eat per appliance, or hardware purchase outright. Targets include financial services, drug discovery, medical data, public sector, and countries without GCP presence. Industry projections cited in the piece estimate 40% of AI training and inference will move outside public cloud by 2027. (Read More)

🫖 TEA For Thought: “More like this will happen. After all, all people care about is privacy and security, and as long as data travels in the cloud, however it’s handled, it’s not safe. Hardware is still the hard block.”

3. 🔍 Perplexity On Why Search Still Matters In The AI Era

Perplexity’s two-stage post-training pipeline for developing high-performance search-augmented language models is outlined in this research article by disentangling deployment constraints from search optimization. The process begins with Supervised Fine-Tuning (SFT) to establish foundational behaviors like guardrails and formatting, followed by Reinforcement Learning (RL) using Group Relative Policy Optimization (GRPO) to enhance factual accuracy and tool-use efficiency. By combining verifiable search-agent QA data with rubric-based general chat data and implementing gated reward aggregation with anchored efficiency penalties, Perplexity successfully improved model performance—achieving higher accuracy and better alignment with user preferences—while significantly reducing operational costs and unnecessary tool usage compared to larger models like GPT-5.4. (Read More)

🫖 TEA For Thought: “Who would have thought that in the age of AI, search is still needed? It’s just in a different way. Instead of us indexing and searching, now it’s all powered by AI, but in the search, accuracy is still very, very important. Perplexity wins not only by the model but also by the harness on top of all of the models. It’s pretty brilliant.”

4. ✂️ The Over-Editing Problem In AI Coding Tools

Researcher nrehiew benchmarked 9 frontier models on a minimal-editing task using 400 programmatically corrupted problems from BigCodeBench. Models were asked to fix single-token bugs (off-by-one errors, flipped operators, boolean swaps), and scored on Pass@1, token-level Levenshtein distance, and added Cognitive Complexity. GPT-5.4 over-edited the most (Levenshtein 0.395 reasoning, 2.31 added complexity) with only 0.723 Pass@1. Claude Opus 4.6 produced the smallest diffs and the highest Pass@1 (0.912 reasoning, Levenshtein 0.060). Reasoning models over-edited more than their non-reasoning counterparts by default, but responded better to explicit prompts to preserve original code. The author then trained Qwen3 4B variants with SFT, rSFT, DPO, and RL: SFT scored perfectly in-domain but collapsed out-of-domain (Pass@1 dropped to 0.458, LiveCodeBench down 14.9%). RL was the only method that generalized cleanly, improving all three metrics with no catastrophic forgetting. LoRA at rank 64 nearly matched full RL, and the recipe scaled to Qwen3 14B (Pass@1 0.833, Levenshtein 0.059). (Read More)

🫖 TEA For Thought: “This is so interesting. When you don’t know what you don’t know and only AI does, I guess that’s the problem. How do we make sure we only make minimal changes, which Opus 4.6 does very well? I guess that’s why they’re so good at coding and it’s engineers’ favorite. There is a reason for it. Over-engineering is definitely a problem, not only wasting tokens but also making it harder for humans to review. Humans have to stay in the loop. However, I wonder how long this can keep going, because when everybody is shipping so fast, what time is even left for humans to review the code? We don’t even have time to review all the plans in detail sometimes.”

5. 🇬🇧 UK Labour Market: No Evidence AI Has Replaced Jobs At Scale

A British Progress report analyzed Annual Population Survey data across 412 UK occupations and found no difference in employment trends between occupations most and least exposed to AI. Office for National Statistics payroll data shows wages in high-exposure occupations have grown more slowly since 2019, but the trend predates ChatGPT and cannot be cleanly attributed to AI. Hours worked in AI-exposed occupations have risen modestly relative to unexposed ones, consistent with augmentation raising demand. Inside high-exposure occupations, programmer and finance analyst roles have continued to grow while administrative and clerical roles have contracted, indicating the same AI exposure can produce augmentation or displacement depending on task structure. Adoption data shows usage is concentrated in a small set of tasks: roughly one-fifth of tasks account for the vast majority of usage. The authors stress these findings do not rule out larger future effects, but plausible near-term predictions should be constrained by current adoption patterns, not projected capabilities. (Read More)

🫖 TEA For Thought: “Very interesting report. If you just look around and talk to folks on the street, you will know how big the gap is between those in AI and those who aren’t. However, is this AI revolution different from electricity? I think it is. AI is revolutionizing all of the games that we play. It’s a knowledge revolution, a productivity revolution. It is how humans work. Human intelligence is redefined by AI, whereas electricity was just about how people work. I would still say there is a difference, even though both are paradigm shifts.”

🛠️ Skill of the Day

openai/privacy-filter — OpenAI’s new bidirectional token-classification model for PII detection and masking. 1.5B parameters (50M active via mixture-of-experts), 128K context, Apache 2.0 license, runs in a browser or laptop. Detects 8 span categories (account numbers, addresses, emails, names, phones, URLs, dates, secrets) in a single forward pass with constrained Viterbi decoding. Fine-tunable via opf train, supports CPU or GPU, pipe-friendly CLI (opf redact), and ships with model weights on Hugging Face.

TEAHEE Moment

Stay sharp, stay informed. See you Sunday!

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 🧠 Two Sign-Bit Flips Can Destroy a 30B Model

Researchers Ido Galil, Moshe Kimhi, and Ran El-Yaniv introduced Deep Neural Lesion (DNL), a data-free attack that identifies and flips a handful of sign bits in stored model weights to cause catastrophic failure. Two flips drop ResNet-50 accuracy from 76.1% to 0.0%, 1 to 2 flips collapse Mask R-CNN and YOLOv8-seg detection, and two targeted flips take Qwen3-30B reasoning from 78% to 0%. The method needs only write access to weights (via firmware exploits, rootkits, DMA attacks, or Rowhammer), no training data, and minimal compute, and it bypasses quantization, pruning, and simple checksumming. The authors show that hardening the top 0.1 to 1% of weights provides practical defense. (Read More)

🫖 TEA For Thought: “The vulnerability is for real. The more powerful AI is, the more fragile it is to being attacked by other AI. Security is going to be the theme when it comes to AI, always.”

2. 🤖 Agent-World: Self-Evolving Training Arena for AI Agents

Renmin University and ByteDance Seed released Agent-World, a self-evolving training arena with over 2,000 environments, 19,000+ validated tools across 20 primary categories, and evaluation on 23 agent benchmarks. The system autonomously mines MCP servers, tool docs, and industrial PRDs to build executable environments, then synthesizes verifiable tasks via graph-based tool dependencies and programmatic Python solutions. A closed-loop trainer pairs multi-environment GRPO reinforcement learning with a diagnostic agent that identifies weak environments and regenerates targeted tasks, co-evolving agent policies with environment difficulty. Trained on Qwen3-8B and 14B backbones, Agent-World outperforms strong proprietary baselines across tool use, software engineering, and deep research benchmarks. (Read More)

🫖 TEA For Thought: “This is super cool, providing scenarios for AI to do actual tasks, and another AI doing supervised learning. This is literally do what I do, not what I say.”

3. 💸 DoorDash Brings Stablecoin Payments to 40+ Countries via Tempo

Tempo announced that DoorDash will let users, dashers, and merchants settle transactions in stablecoins across more than 40 countries, citing faster payouts, lower cross-border costs, and transaction flexibility. Co-founder Andy Wang framed faster, cheaper merchant and dasher payouts as a “no-brainer for the entire ecosystem.” The integration lands alongside Tempo rollouts with Stripe, Paradigm, Coastal Bank, and ARQ. DoorDash delivered 903 million orders worth $29.7 billion in Q4 2025, making this one of the largest consumer apps to adopt a stablecoin payment rail for everyday settlements. Stripe, Visa, and Mastercard have all expanded stablecoin infrastructure via recent deals (Bridge, BVNK, and settlement platform expansions). (Read More)

🫖 TEA For Thought: “DoorDash might just open the doors for Tempo via stablecoins.”

4. 🦉 Duolingo Makes B2-Level Content Free in Nine Languages

Duolingo announced free access to B2-level (CEFR) advanced learning content across English, Spanish, French, German, Italian, Portuguese, Japanese, Korean, and Chinese, available on web, iOS, and Android. Previously capped at A2 or B1, the free tier now includes Advanced Stories for reading comprehension and DuoRadio for listening. Competitors like Babbel and Busuu keep advanced content behind paywalls. Duolingo reports 52.7 million daily active users (up 30% year over year) versus 12.2 million paid subscribers, and it cites ACTFL research showing a second language can raise employability by up to 50%. The move doubles as a growth play after Duolingo projected softer Q2 2026 bookings. (Read More)

🫖 TEA For Thought: “It’s a good model switch for Duolingo. Now that knowledge is free with AI, why would you pay to access premium content when you can get it for free?”

5. 📚 Agentivism: A Learning Theory for the AI Age

Lixiang Yan and Dragan Gašević propose Agentivism, a new learning theory arguing that existing frameworks (behaviourism, cognitivism, constructivism, connectivism) cannot fully explain learning when generative and agentic AI lets learners delegate explanation, writing, and problem solving to systems that act on their behalf. The authors contend that successful task performance can no longer be assumed to indicate learning, since AI-assisted users may complete work effectively while developing weaker understanding, judgment, and transfer. Agentivism defines learning as durable growth in human capability through selective delegation to AI, epistemic monitoring and verification of AI outputs, reconstructive internalization, and transfer under reduced AI support. (Read More)

🫖 TEA For Thought: “How learning is defined, how humans learn, and what humans learn when things can be done for us. A paradigm shift, for sure.”

🛠️ Tools of the Day

koala73/worldmonitor — Real-time global intelligence dashboard with AI news aggregation. 51.3K stars.

KeygraphHQ/shannon — Autonomous white-box AI pentester for web apps and APIs. 39.4K stars (+346 today).

BasedHardware/omi — Ambient AI that watches your screen and offers guidance. 11.9K stars (+3.9K this week).

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 💸 Microsoft Moves GitHub Copilot to Token-Based Billing

Leaked internal documents reported by Where’s Your Ed At reveal that Microsoft plans to pause new signups for the student and paid individual tiers of GitHub Copilot, tighten rate limits, and eventually shift users to token-based billing. The weekly cost of running Copilot has nearly doubled since January, which the documents frame as the reason the transition became urgent. Today, Pro users ($10/month) get 300 “requests” and Pro+ users ($39/month) get 1,500, with pricier models costing more requests. Under token-based billing, users would instead pay for actual token consumption, where Claude Opus 4.7 alone costs $5 per million input tokens and $25 per million output tokens. Microsoft has now confirmed parts of the plan in an April 20 GitHub blog post, and the shift follows similar moves by Anthropic, OpenAI, and Cursor as the era of subsidized AI pricing ends. (Read More)

🫖 TEA For Thought: “Well, everyone is using it.”

2. 🎵 Deezer: 44% of Daily Uploads Are AI-Generated

Deezer announced on Monday that AI-generated tracks now make up 44% of all new music uploaded to its platform, with the service receiving almost 75,000 AI tracks per day and over 2 million per month. The ramp has been steep: 10,000 per day in January 2025, 30,000 in September, 50,000 in November, 60,000 in January 2026, and 75,000 now. Actual consumption remains low at 1-3% of total streams, and Deezer says 85% of those streams are flagged as fraudulent and demonetized. The platform will also stop storing hi-res versions of AI tracks. The numbers land the same week an AI-generated song topped the iTunes charts in the US, UK, France, Canada, and New Zealand, and on the back of a November Deezer/Ipsos survey in which 97% of participants could not distinguish fully AI-generated music from human-made music. (Read More)

🫖 TEA For Thought: “The thing is, when there is so much AI-generated art but no consumer to appreciate it...”

3. 🗣️ Iowa State Study: News Writers Rarely Anthropomorphize AI

An Iowa State University study published in Technical Communication Quarterly analyzed the News on the Web (NOW) corpus of over 20 billion words from English-language news across 20 countries and found that journalists pair AI terms with mental verbs like “think,” “know,” and “understand” far less often than everyday speech does. The most common pairing, “AI needs,” appeared 661 times, while “ChatGPT knows” appeared only 32 times. Researchers Jo Mackiewicz, Jeanine Aune, Matthew J. Baker, and Jordan Smith argue that AP editorial guidelines, which discourage attributing human emotions to AI, likely shape this restraint. They also find anthropomorphism exists on a spectrum: “AI needs data” is mechanical, while “AI needs to understand the real world” edges closer to implying human reasoning. The team warns that even subtle mental-verb framing can overstate AI capability and distract from the developers and organizations actually responsible for the systems. (Read More)

🫖 TEA For Thought: “Language matters.”

4. 🏢 Steve Yegge: Inside Google’s Two-Tier AI Adoption

In an April 20 follow-up thread, Steve Yegge says multiple Googlers from different orgs reached out anonymously after his earlier post on Google’s AI adoption, corroborating and sharpening his original picture. They describe a two-tier system where DeepMind engineers use Claude as a daily tool while most of the rest of Google does not. When internal talk turned to equalizing access, the proposed fix was reportedly to remove Claude for everyone, which DeepMind pushed back on so hard that several engineers threatened to leave. Non-DeepMind engineers get routed to internal Gemini variants behind router-style names that obscure the underlying model, with reports of regressions severe enough that some senior staff stopped using the tools. Leadership has responded by mandating AI use in OKRs and standing up an internal token-usage leaderboard, while telling managers simultaneously that it will and will not factor into performance reviews. Yegge closes: “Nobody is as far ahead as they might look from the outside, and none of you are as far behind as you might be worried you are.” (Read More)

🫖 TEA For Thought: “Nobody is as far ahead as they might look from the outside, and none of you are as far behind as you might be worried you are.”

5. 🔋 EU Mandates User-Swappable Phone Batteries by February 2027

Android Headlines reports that, starting February 2027, smartphones sold in the EU must be designed so users can replace the battery themselves using basic tools, not specialized equipment. Batteries must retain at least 80% capacity after 800 charge cycles, repair manuals must be publicly available, and OEMs must keep spare parts in circulation for years. The rule builds on phase one, already in effect since June 2025, which mandates up to 10 years of repair support, stricter durability benchmarks, and a ban on blocking independent repair services. The new mandate stops short of pop-off back covers and sits between the removable batteries of the past and today’s glued-in designs. Because OEMs are unlikely to produce two hardware versions of the same device, the regulation is expected to reshape smartphone design worldwide, not just in the EU. (Read More)

🫖 TEA For Thought: “The EU’s framework is just crazy...”

🛠️ Tools of the Day

sansan0/TrendRadar. AI-powered public opinion monitor with multi-source intelligent alerts. 53.5K stars on GitHub Trending.

zilliztech/claude-context. Full codebase context server for Claude Code and other coding agents. 6.4K stars on GitHub Trending.

openai/openai-agents-python. Lightweight multi-agent workflow framework from OpenAI. 24.3K stars on GitHub Trending.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. ⏳ The 12-Month Window: When AI Founders Should Sell

On a recent No Priors episode, AI investor Elad Gil told co-host Sarah Guo that most companies have roughly a 12-month peak window before value crashes, and the firms that capture generational returns are the ones that catch that moment instead of assuming good times will keep compounding. He cited Lotus, AOL, and Mark Cuban’s Broadcast.com as classic cases of pulling the ripcord at the top. Gil’s practical tip for founders: pre-schedule a board meeting once or twice a year dedicated specifically to discussing exits, which drains emotion out of the decision. He framed this as especially urgent for AI startups, many of which exist only because foundation models have not yet expanded into their category, a gap that Deel CEO Alex Bouaziz and others have openly acknowledged will not last. Gil’s framing question: “Are these next six months when I’m going to be the most valuable I’ll ever be?” (Read More)

🫖 TEA For Thought: “The gap is only getting wider.”

2. ⌚ Fitbit Air: Google Rebrands Health Stack Around the “G”

9to5Google reports that Google’s upcoming screen-less health band, teased at the end of March and already worn by Stephen Curry, will launch as the “Google Fitbit Air.” The “Air” name nods to the wearable’s thinner, all-day-wear design, echoing the 2019 Aria Air scale. More significantly, the Fitbit Premium subscription is being rebranded as “Google Health,” and the personal health coach currently in public preview will be called “Google Health Coach.” Through the testing period, the software has been branded only as “Coach,” with no Fitbit tag. The move pulls wellness features into the core Google brand rather than Fitbit, and it explains why Curry’s teaser ended on a gradient “G” with no Fitbit logo. An official announcement is expected in the coming weeks. (Read More)

🫖 TEA For Thought: “Hardware is the hard problem for big tech. The competition will only get more severe.”

3. 🔎 Exa’s Canon: Composing a Search Engine as a DAG

Exa published a technical post on Canon, the search pipeline orchestrator it built to run billions of search requests for AI agents. The team defines the entire search pipeline as a DAG of typed nodes (retrieve, classify, localize, rank), so the executor can fire each node concurrently the moment its inputs are ready. A single query in 250ms runs classification, localization, retrieval, and ranking across multiple indexes in parallel, with every node observable for debugging. Exa argues this composable approach is essential now that most pipeline code is written by agents: agents produce locally correct code but struggle with global constraints, so the DAG contract forces each node to be independently reasoned about. The post details how a simple query has grown into a graph of 20-plus node types serving thousands of AI agents with per-customer branches. (Read More)

🫖 TEA For Thought: “Isn’t this just looking more and more like blockchain? The convergence is happening.”

4. ☠️ Nvidia’s China Share Falls Below 60% as Domestic Chips Surge

IDC data reported by Reuters shows Chinese semiconductor firms captured 41% of the domestic AI server market in 2025, delivering 1.65 million AI GPUs out of roughly 4 million total units. Nvidia still leads at 55% (about 2.2 million cards), a sharp contraction from its pre-sanctions 95% share. Huawei is the biggest domestic winner with 812,000 chips shipped (nearly 20% of the market) and just launched its Atlas 350 accelerator, which it claims delivers roughly three times the performance of Nvidia’s H20. Alibaba’s T-Head placed third at 256,000 units, AMD shipped 160,000 (4%), and Baidu’s Kunlunxin and Cambricon rounded out the top five at 116,000 each. Trump reversed the H20 and MI308 ban in July 2025 and allowed H200 shipments to China in December 2025, but Beijing is actively pushing data centers toward domestic chips to close the five-to-ten-year gap in AI data center silicon. (Read More)

🫖 TEA For Thought: “China is really speeding up the full stack of AI: chips, infra, cloud, models, and apps. The gap is closer than ever.”

5. 🧠 Memory Transfer Learning: Cross-Domain Memory for Coding Agents

A new arXiv paper by Kangsan Kim, Minki Kang, Taeil Kim, Yanlai Yang, Mengye Ren, and Sung Ju Hwang (submitted April 15, 2026) introduces Memory Transfer Learning (MTL), a framework for coding agents that pools memory across heterogeneous task domains instead of restricting it to a single domain silo. The authors evaluate four memory representations (from concrete execution traces to abstract insights) across six coding benchmarks and find that cross-domain memory improves average performance by 3.7%, primarily by transferring meta-knowledge such as validation routines rather than task-specific code. A key finding: abstraction dictates transferability. High-level insights generalize well, while low-level traces often cause negative transfer due to excessive specificity. Transfer effectiveness also scales with memory pool size, and memory can move between different models. Project page at memorytransfer.github.io. (Read More)

🫖 TEA For Thought: “The shared memory between all agents in any org, consolidated memory, dreaming, is what everyone’s working on.”

🛠️ Tools of the Day

lsdefine/GenericAgent — Self-evolving agent that achieves system-level control with reduced token usage. ~3.5K stars this week, GitHub trending.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 👁️ Zoom Teams Up With Sam Altman’s World To Verify Humans In Meetings

Zoom announced a partnership with World, Sam Altman’s human ID verification company, to confirm meeting participants are real people and not AI-generated deepfakes. The feature uses World ID Deep Face, which cross-references a signed image from the user’s registration through World’s Orb device, a real-time face scan from the user’s device, and a live video frame, only granting a “Verified Human” badge when all three match. Hosts can enable a Deep Face waiting room to require verification, and participants can request mid-call that someone verify themselves. The threat is real: engineering firm Arup lost $25 million in early 2024 after an employee authorized wires on a video call where every other participant was an AI deepfake, and deepfake-enabled fraud topped $200 million in Q1 2025 with average corporate losses above $500,000 per incident. World has also partnered with Tinder, Visa, and AI shopping agent verification. (Read More)

🫖 TEA For Thought: “There’s more control in the name of protecting you. This is Zoom just revealing its true color.”

2. ☠️Circle CEO: China Could Launch Yuan Stablecoin Within 3 To 5 Years

Circle CEO Jeremy Allaire told Reuters in Hong Kong there is a “tremendous opportunity” for a yuan-backed stablecoin, predicting China could roll one out within three to five years as digital currencies become more integrated into global trade. Reuters reported in August 2025 that Chinese officials were already exploring a yuan stablecoin to boost RMB international adoption, a notable turn given China’s ban on crypto trading and mining since 2021. Allaire has pushed this argument since 2023, framing stablecoins as a superior vehicle for RMB internationalization over CBDCs. Experts note a true yuan stablecoin would require Beijing to make the RMB fully convertible, meaning capital controls, not technology, remain the real blocker. An offshore yuan (CNH) stablecoin fits existing controls, while an onshore yuan (CNY) version does not. The global stablecoin market sits near $315 billion today, dominated by USDT and USDC. (Read More)

🫖 TEA For Thought: “Stablecoin is just a CBDC in disguise.”

3. 🧠 Hyperliquid’s Ecosystem: The House Of All Finance

Pink Brains published a deep thread mapping Hyperliquid’s ecosystem as the userland extension of a functioning exchange, arguing HyperEVM only makes sense for financial primitives and trading infra, not GameFi or SocialFi. The DEX layer includes prjx_hl (86% of fees to LPs, nearly $300K daily fees in month one), HyperSwapX, HybraFinance, and Kittenswap forks of Velodrome with stablecoin fees as low as 0.01-0.02%. Lending is anchored by Felix Protocol on Morpho ($462M deposits, $124M borrows) and HyperLend, an Aave V3-style fork. Earn includes Kinetiq, the dominant HYPE LST with $956M+ TVL and 82.5% staking market share. HIP-3 lets any team with 500K HYPE staked deploy perps on pre-IPO equities, commodities, or indices: tradexyz drives ~90% of HIP-3 volume, and ventuals offers SpaceX/OpenAI/Anthropic pre-IPO perps. HIP-4 will enable cross-margined prediction market contracts against ETH perps in the same account. (Read More)

🫖 TEA For Thought: “Sometimes it’s hard to imagine this is developed by a team of 11.”

4. 🪪 Anthropic Now Asks For Your Passport And Selfie To Use Claude

Anthropic quietly published identity verification requirements on April 14, asking certain Claude users to submit a government-issued photo ID and a live selfie, something no other major AI chatbot currently requires. The company uses Persona Identities (the same KYC infrastructure used across financial services) and requires a physical passport, driver’s license, or national identity card. Photocopies, mobile IDs, and student credentials do not qualify. Verification triggers during “certain capabilities,” “routine platform integrity checks,” or safety and compliance measures. Anthropic says data stays on Persona’s servers, is encrypted in transit and at rest, excluded from model training, and not shared for marketing. The move comes weeks after millions fled OpenAI for Anthropic over the Pentagon surveillance deal, with daily signups breaking records and free users up 60% since January. Chinese users accessing Claude through intermediaries are most affected, since a live selfie matched against a government document is hard to bypass. An October 2025 Discord breach exposed roughly 70,000 government IDs submitted for age verification, showing third-party ID custody carries real risk. (Read More)

🫖 TEA For Thought: “Anthropic’s god complex. This is all about control. Private AI is the future, and open-sourced models deployed locally are the future. The battle between proprietary and open-sourced models is just the beginning.”

5. 💳 Four Models For Enterprise Credit Governance (OpenAI, Cursor, Clay, Vercel)

The Financial Engineer test-drove credit governance architectures at four AI-native companies and found each answers the enterprise hierarchy-and-enforcement problem differently. OpenAI uses a “decision waterfall” where every API request passes through one evaluation path that checks rate limits, credits, and entitlements synchronously, with a three-level Organization to Projects to API Keys hierarchy and usage tiers from Free ($100/month cap) to Tier 5 ($50,000/month cap). Cursor migrated in June 2025 from 500 fast requests per month to dollar-denominated credit pools ($20 on Pro, ~$400 on Ultra) consumed at raw LLM API cost with zero markup, and transitions from per-user pools on Teams to pooled credits on Enterprise. Clay runs a dual-currency model separating Actions (platform orchestration, 1 per enrichment) from Data Credits (marketplace data, 0.5 to 10+ per lookup), with unlimited users on every plan and per-workbook spend caps. Vercel skips deep hierarchy modeling and invests in observability: the AI Gateway’s Custom Reporting API exposes spend across six groupBy dimensions (day, user, model, tag, provider, credential type) with request-time attribution tags, plus spend management webhooks that can pause project deployments at threshold. (Read More)

🫖 TEA For Thought: “Very interesting, especially when companies get bigger. How do you deal with token allocation? Something to think about and research.”

🛠️ Tools of the Day

OpenBMB/VoxCPM: Tokenizer-free TTS for multilingual speech generation, creative voice design, and true-to-life cloning. 14.1k stars.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

When people talk about becoming an “AI-first” company, the immediate instinct is to incorporate AI into existing workflows, typically starting with basic automation. However, as Peter describes, and as Jack Dorsey has similarly noted, this is not merely about adding automation to what already exists. It is about a deeper structural shift, rethinking how a company is organized and how the business operates from the ground up.

The Self-Healing Brain and Entangled Software

The most critical component of this shift is the self-learning and self-healing layer: the true “brain” of the AI. This layer constantly learns from memory sessions, recognizes patterns, and proactively heals systems before they break. This continuous evolution is the essential part of an AI-native architecture.

In the realm of software sales, particularly for SaaS and service businesses, the CEO of CrewAI introduced a compelling concept: “entangled software” . In this model, customer behavior dictates how the software should function, and simultaneously, the software’s functionality influences how customers behave. It is a continuous loop of entanglement where a change at any endpoint dynamically influences the other.

Currently, this proactive, self-evolving layer seems to be the missing link that the industry is racing to build. As Peter Steinberg recently mentioned in an interview, if he had free time, he would focus on the concept of AI “dreaming”: the nighttime process of consolidating memories and logs to improve and self-evolve, much like the human brain.

The Road to the Ideal State

Right now, much of the industry’s attention is concentrated on agent frameworks and harnesses, especially following the open-source release by Claude Code. While these tools are important, we must not lose sight of the end goal: solving problems.

To solve problems, we must first identify them, and AI excels at this. The missing layer is the intelligence that learns, spots patterns, adapts, and becomes proactive. As the CEO of CrewAI aptly stated, you can build a road, but you must know where it leads. The destination is the “ideal state” where the problem is resolved, and the self-learning, self-healing aspects of AI are the vehicles that will get us there.

While it may sound like we are envisioning a utopian future where this ideal state is reached without human involvement, the reality is that we are not there yet. We are still struggling to build this self-healing, self-evolving intelligence. Reaching that point will require significant thought and structural innovation.

The Future of Thought: Coach or Crutch?

Looking ahead, consider the toddlers of today and what their future will look like. Will they need to use their brains as much when so many problems are already solved by AI? A thought-provoking discussion (The Last Generation of Humans Who Will Ever Have to Think) on X highlighted how dystopian this future could potentially become, posing a critical question: Will we use AI as a coach or as a crutch?

If we use AI as a coach, it becomes a 24/7 personalized tutor. It is patient, available, asks challenging questions, and acts as a thought partner, ultimately helping us learn and grow.

However, for many, it is not an “if” but a “when.” As AI becomes increasingly powerful, the necessity for human thought diminishes. We are already seeing students rely heavily on AI tools because it removes the burden of thinking, causing our mental “muscles” to shrink. If we allow AI to make all our decisions, we are essentially surrendering control of our lives. Given current trends, such as Claude tracking IDs and restricted access, it is easy to feel that we are already heading down this path.

The Weightlifting of the Mind

What, then, does that future hold? Thinking is like weightlifting: it requires muscle, training, and consistent effort. The more you think, the stronger you become. Perhaps this ability to think, even when AI can do it for us, is what fundamentally makes us human.

If we are moving toward a world where there is seemingly nothing left for us to think about, it raises a profound question: In a critical moment, what will we choose to do, and how will we choose to think?

Till then, cheers!

Subscribe now

Previous column articles can be found here:

https://newsletter.ownlyagent.com/t/teacolumn

1. 🤖 Ukraine’s Ground Robots Now Run 9,000 Missions a Month

Ukrainian President Volodymyr Zelenskyy claimed that ground robots and drones, acting on their own, forced a group of Russian soldiers to surrender a fortified position, an incident that if verified would mark a robotic milestone in the war. A promotional video from Zelenskyy credited Ukraine’s military robots with over 22,000 missions in the last three months, while Ukraine’s defense ministry reported a threefold increase in uncrewed ground vehicle missions over the last five months, with more than 9,000 robotic missions run in March alone. The episode likely traces back to a Kharkiv Oblast operation by the 3rd Separate Assault Brigade, which used flying drones and “kamikaze” ground robots against Russian frontlines. Persistent drone surveillance has created a 12-mile (20 km) “kill zone” beyond the frontline as of February 2026, with drones now causing the majority of battlefield casualties on both sides as the full-scale war enters its fifth year. The newest Ukrainian drones integrate onboard AI so they can track and strike targets even when jammed. (Read More)

🫖 TEA For Thought: “This is definitely revolutionizing how people think about war. You do not need big tanks, you do not need expensive warships. If you have technology as cheap as drones, it is another level.”

2. 🔗 ChatGPT Cites Half the URLs It Retrieves, and Reddit Is 67.8% of the Rest

Ahrefs analyzed 1.4 million ChatGPT prompts from February 2025 covering 47.1 million retrieved URLs and found ChatGPT cites only 49.98% of them. The citation rate varies sharply by source: Search results hit 88.46% (25.6M data points), News 12.01%, Reddit just 1.93% despite 16.2M retrievals, YouTube 0.51%, and Academia 0.40%. Reddit alone accounts for 67.8% of the entire non-cited pool, meaning ChatGPT pulls Reddit at scale but almost never credits it. On average, ChatGPT retrieves 16.57 cited and 16.58 non-cited URLs per prompt. Cited pages have a median age of about 500 days (1.3 years), while non-cited pages are overwhelmingly very young. Title relevance matters: cosine similarity between prompt and cited URL title averages 0.602 versus 0.484 for non-cited, and natural-language URL slugs get cited 89.78% of the time versus 81.11% for opaque slugs. (Read More)

🫖 TEA For Thought: “I guess almost 70% of the content is from Reddit. That is why Reddit’s CEO is so harsh on AI agent control.”

3. ⚖️ Federal Judge Rules AI Chats Are Not Privileged, Lawyers Scramble

A federal judge in New York ruled that Bradley Heppner, former chair of bankrupt financial services company GWG Holdings and founder of Beneficent, cannot shield his AI chatbot conversations from prosecutors pursuing securities and wire fraud charges filed last November (Heppner pleaded not guilty). The court found the chats failed basic privilege tests: they were not communications with a lawyer, the user could not reasonably expect confidentiality, and the work was not performed at the direction of counsel. White-collar defense firm Sher Tremonte added new contract language in March warning that “disclosure of privileged communications to a third-party AI platform may constitute a waiver of the attorney-client privilege.” Debevoise & Plimpton advised clients that if counsel does authorize AI use, the prompt itself should say so, suggesting language like “I am doing this research at the direction of counsel for [X] litigation.” Prosecutors and civil litigants can now demand Claude and ChatGPT logs in discovery. (Read More)

🫖 TEA For Thought: “Thought crime might be an actual crime in the eyes of AI chatbots.”

4. 💸 X Rolls Out Smart Cashtags With Live Prices and Wealthsimple Trading

X launched Smart Cashtags in April 2026, a feature that turns crypto tickers and stock symbols inside posts into tappable links showing live price charts and asset-specific feeds without leaving the platform. The rollout is initially available to iPhone users in the United States and Canada, with a pilot trading integration via Wealthsimple for Canadian users. X head of product Nikita Bier called it “just the first step” in positioning X as a destination for finance and crypto activity. The feature pairs with X Money, which is entering early public access in April 2026 with planned peer-to-peer transfers, bank deposits, and debit-card functionality, with Visa backing parts of the payments infrastructure. Bitcoin traded at $74,662 at publication, and the launch leans into X’s existing role as the place traders watch narratives and sentiment move in real time. (Read More)

🫖 TEA For Thought: “You are going to see how AI agents get plugged into this place.”

5. 🛒 Coinbase Launches Bazaar MCP, the First Discovery Marketplace for AI Tools

Coinbase Developer Platform launched Bazaar MCP, billed as the first discovery marketplace where AI agents can search for a capability, check the price, pay the micro-transaction, and use the tool without any human hand-holding or API keys. Built on x402, Coinbase’s HTTP-native payments protocol that turns the dormant HTTP 402 status code into a real payment layer using stablecoins like USDC, Bazaar ranks tools by real usage and trust so agents route to the best services. Providers list an x402 endpoint with schema, price, and description, and any agent on the network can discover and call it permissionlessly. The announcement builds on Coinbase’s AgentKit (2,000-plus developers, 20,000-plus agents deployed, 600,000-plus transactions on Base and Base Sepolia) and Agentic Wallets, which give agents programmable spending policies, non-custodial identity, and gasless transactions on Base. (Read More)

🫖 TEA For Thought: “Base is really building the network for authentic payments.”

🛠️ Tools of the Day

google/adk-python — Google’s open-source, code-first Python toolkit for building, evaluating, and deploying multi-agent systems, MCP-first with Apache 2.0 license. 19,022 stars.

TEAHEE Moment

Stay sharp, stay informed. See you Sunday!

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. ⭐ 6 Million GitHub Stars Were Fake, and VCs Funded Some of Those Repos

A Carnegie Mellon study found approximately 6 million suspected fake stars across 18,617 GitHub repositories, sold openly for $0.03 to $0.85 each. The perverse logic: VCs use star counts as funding signals (median seed-stage: 2,850 stars), so a $285 purchase can manufacture credibility worth millions. Detection is possible. Fake campaigns leave fingerprints: zero-follower accounts, near-zero fork-to-star ratios, and sudden spike patterns. One 157,000-star repo had only 17 forks per 1,000 stars. Flask has 235. The FTC’s 2024 rule bans fake social metrics at $53K per violation. (Read More)

🫖 TEA For Thought: “100% — wherever there are people and competition, there will be exploitation.”

2. 🔐 AI Security Is Now a Token Budget War

The UK AI Safety Institute confirmed it: spending more tokens on defense yields better vulnerability detection. The formula is simple. To harden a system, you need to spend more tokens discovering exploits than attackers will spend attacking it. This creates an interesting inversion for open source. Security investments on shared codebases benefit all users simultaneously, making collaborative defense more cost-efficient than individual reimplementation. AI-driven security is a computational arms race where resources determine defensive strength. (Read More)

🫖 TEA For Thought: “This is proof of work. With tokens, backed by real money. The best defense can always be discovered. That is the ‘crypto’ of cybersecurity.”

3. 💸 Anthropic Ends the Flat-Fee Era

Anthropic restructured enterprise pricing. Seat fees now cover platform access only. Claude, Claude Code, and Cowork usage is billed separately at standard API rates. The trigger: GPU rental costs climbed 48%, CoreWeave raised rates 20%, and API uptime dropped to 98.95% over 90 days (well below the 99.99% cloud standard). OpenAI, GitHub, and Windsurf have already made the same move. The subsidized-access era for heavy users is over industry-wide. (Read More)

🫖 TEA For Thought: “The seat charge era might be over soon. Folks will be charged for the tokens they actually consume. More people will direct their work toward open-source, locally deployed models. Get some hardware ready before prices go even further up.”

4. ⚛️ NVIDIA Just Built the AI Brain for Quantum Computers

NVIDIA launched Ising, the world’s first family of open-source quantum AI models. The family has two components. Ising Calibration is a vision language model that automates quantum processor calibration, cutting setup time from days to hours. Ising Decoding is a neural network optimized for quantum error correction, delivering 2.5x faster performance and 3x higher accuracy than pyMatching, the current industry standard. Harvard, Fermilab, Lawrence Berkeley Lab, and IQM Quantum Computers are already adopting it. Jensen Huang framed it plainly: “AI becomes the control plane — the operating system of quantum machines.” The quantum sector is projected to exceed $11 billion by 2030. (Read More)

🫖 TEA For Thought: “Quantum plus AI = unstoppable.”

5. 🤖 One Markdown File. 235 GPU Kernels. 38% Faster.

Cursor ran a multi-agent system against 235 CUDA kernel optimization problems for NVIDIA’s Blackwell B200 GPUs. A single planner agent coordinated autonomous workers. The only instruction: one markdown file defining the output format, rules, and test criteria. Result: 63% of kernels outperformed the baseline, with a 1.38x geomean speedup. 19% exceeded 2x improvement. Work that traditionally takes kernel engineers months or years. The key insight: clear objective definition, not complex orchestration. (Read More)

🫖 TEA For Thought: “This is amazing. If you can define the ideal state as accurately as possible, AI agents can figure out the best way to achieve it.”

🛠️ Tools of the Day

NousResearch/hermes-agent. “The agent that grows with you.” Adaptive agent framework that builds on prior context across sessions. +53K stars this week.

HKUDS/DeepTutor. Agent-native personalized learning assistant. Adapts to the learner over time instead of serving static content. +5.5K stars this week.

HKUDS/AI-Trader. 100% fully-automated agent-native trading. Multi-agent team handling research, analysis, and execution. +800 stars this week.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 🌐 Google Adds AI Skills to Chrome

Google launched “Skills” in Chrome: save AI prompts as reusable workflows, trigger them with / or + on any tab. Built on Gemini, Skills let you automate recurring tasks like macro calculations, shopping comparisons, or document scanning across tabs. A pre-built Skills Library ships for productivity, recipes, and budgeting. Rolling out today to desktop users in English (US). (Read More)

🫖 TEA For Thought: “Google has been on fire lately. This is also something that Perplexity did about half a year ago.”

2. 🔬 AI Agents Still Flunk Real Science, But the Gap Is Closing Fast

Ai2’s benchmarks ScienceWorld and DiscoveryWorld test whether AI can do science, not just answer science questions. In 2022, top models scored below 10% on elementary-school experiments. Today they hit the low 80s on ScienceWorld. But on DiscoveryWorld’s open-ended discovery tasks, frontier agents complete only ~20% at higher difficulty while human scientists solve ~70%. The gap: knowing the boiling point of water versus figuring out how to measure it yourself. (Read More)

🫖 TEA For Thought: “What’s scary is when book-smart agents finally get street smarts. They become unstoppable. And that day is coming sooner than anyone expects.”

3. 🏗️ Let AI Be Brilliant, Let Code Be Accurate: The Hybrid Workflow Pattern

Will Larson documents a production-proven pattern: prototype with full agent control, then systematically replace agent decision-making with deterministic code. His security alert system shows it in action: a script filters webhook severity and packages metadata, then hands off only owner-identification and message formatting to agents. Result: 100% reliable routing. Pure agent workflows failed despite repeated prompting. The formula: code handles flow control, agents handle ambiguity. (Read More)

🫖 TEA For Thought: “It is very important to let AI do what it does best, but also to keep the deterministic layer so that things can be executed accurately.”

4. 📊 74% of AI’s Value Goes to 20% of Companies, and the Gap Is Widening

PwC surveyed 1,217 senior executives across 25 sectors. Finding: 74% of AI’s economic gains are captured by just 20% of organizations. The leaders are not deploying more tools. They are using AI to reinvent business models and pursue growth from industry convergence. They are 2.8x more likely to increase decisions made without human intervention, 2.6x more likely to report AI enables business model reinvention, and 1.7x more likely to have a Responsible AI framework. The divide is structural, not technological. (Read More)

🫖 TEA For Thought: “This is the inflection point. The gap between those who use AI and those who don’t is going to grow wider and wider.”

5. ⛏️ 20 Parallel Agents, 1,039 Experiments, 1st Place: Proof-of-Work for AI

Ryan Li won the Paradigm Autoresearch hackathon by running 20 parallel Claude Code agents simultaneously, each sweeping different strategy spaces on a prediction market challenge. No human-designed strategy. Agents saved learnings to shared markdown so each could build on the others’ work. The biggest breakthrough came when one agent ignored all existing code and started fresh, discovering a superior architecture. Final strategy validated across 3,200 simulations across 16 seeds. Result: 1st place with $42.32 edge, beating competitors who overfit to a single seed. (Read More)

🫖 TEA For Thought: “This is like blockchain-style proof of work for agents. Every agent runs its own method until it finds the best solution. The reward is not money but the winning strategy itself. Very interesting.”

🛠️ Tools of the Day

microsoft/markitdown. Converts any file (PDF, DOCX, PPTX, XLSX, images) to clean Markdown. The quiet workhorse every agent pipeline should have at the front door. +15K stars this week.

thedotmack/claude-mem. Auto-captures everything Claude Code does during a session and compresses it into persistent memory. Removes the “start from zero every time” tax. +8.7K stars this week.

virattt/ai-hedge-fund. Multi-agent team simulating a hedge fund (researcher, analyst, risk manager, portfolio manager). Reference architecture for building specialized agent swarms. +3.4K stars this week.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 🏭 Maine Moves to Pause New AI Data Centers Until 2027

Maine’s legislature approved bill text in both chambers that would pause new AI data center construction in the state until November 1, 2027, with a final vote expected by April 15. If signed, it would be the first statewide moratorium in the US. Business Insider tracked 12 similar state bills introduced in 2026 across Georgia, Maryland, Michigan, New Hampshire, New York, Oklahoma, South Carolina, South Dakota, Vermont, Virginia, Wisconsin, and Maine, with only Maine’s advancing. The US currently has 4,000 data centers with 3,000 more proposed or under construction per the American Edge Project. Virginia, home to the world’s largest data center concentration, punted its moratorium to 2027, and its sales tax exemption cost the state $1.9 billion in fiscal 2025. Sen. Bernie Sanders and Rep. Alexandria Ocasio-Cortez introduced a federal moratorium bill in March. Local moratoriums have also passed in East Lansing, Tulsa, and Port Washington, Wisconsin. (Read More)

🫖 TEA For Thought: “If this trend continues, and it will, data centers will have to find the next paradigm. The hardware-limited GPU and TPU might become obsolete. The ones not limited by physical hardware might be the next paradigm.”

2. ⚡ Tufts’ Neuro-Symbolic AI Cuts Training Energy to 1% of Standard Models

Tufts University researchers led by Professor Matthias Scheutz unveiled a neuro-symbolic AI system that combines neural networks with symbolic reasoning, cutting training energy to 1% of standard visual-language-action (VLA) systems and operational energy to 5%. On the Tower of Hanoi puzzle, the hybrid system hit 95% success versus 34% for conventional VLAs, and 78% on an unseen harder variant where traditional models failed every attempt. Training dropped from more than a day and a half to 34 minutes. The work will appear at the International Conference of Robotics and Automation in Vienna in May and is posted on arXiv (Duggan, Lorang, Lu, Scheutz, Feb 22 2026). Context: the IEA reports AI systems and data centers used 415 terawatt hours in 2024, over 10% of US electricity, with demand projected to double by 2030. Scheutz noted a Google AI summary can consume up to 100 times more energy than generating the underlying website listings. (Read More)

🫖 TEA For Thought: “This is the flywheel effect. The more advanced the tech is, the more energy it can save using its own energy.”

3. 🧠 LangChain: Your Harness Is Your Memory

LangChain’s Harrison Chase argues that agent harnesses (Claude Code, Deep Agents, Pi/OpenClaw, OpenCode, Codex, Letta Code) are now the dominant way to build agents and are tightly coupled to memory. Citing the leaked 512k lines of Claude Code source as evidence of how much harness code exists, he warns that closed harnesses behind proprietary APIs (Anthropic’s Claude Managed Agents, OpenAI’s Responses API with server-side compaction, Codex’s encrypted compaction summaries) create platform lock-in by owning long-term memory. Short-term memory (conversation, tool results) and long-term cross-session memory both live inside the harness, covering CLAUDE.md loading, skill metadata, compaction rules, and filesystem exposure. Chase shares an anecdote about his accidentally-deleted Fleet email assistant requiring full re-teaching of tone and preferences. LangChain’s answer is Deep Agents: open source, model-agnostic, uses agents.md and skills standards, with Mongo, Postgres, and Redis plugins for memory storage and self-hostable deployment via LangSmith. (Read More)

🫖 TEA For Thought: “Security is more critical and crucial than ever.”

4. 🧱 Cursor, Claude Code, and Codex Are Merging Into One AI Coding Stack

Cursor, Claude Code, and OpenAI Codex are assembling into a composable coding stack with three distinct layers rather than consolidating into one tool. On April 2, Cursor launched version 3 with a dedicated Agents Window for running parallel agents across local machines, worktrees, and cloud sandboxes, plus a /best-of-n command that sends the same prompt to multiple models. Three days earlier, OpenAI published codex-plugin-cc on GitHub, an Apache 2.0 plugin that runs Codex directly inside Anthropic’s Claude Code with six slash commands including /codex:adversarial-review. According to a Pragmatic Engineer survey of 906 engineers, Claude Code holds a 46% most-loved rating, and SemiAnalysis estimates it accounts for roughly 4% of public GitHub commits. Codex passed 3 million weekly active users in March 2026. The stack splits into orchestration (Cursor), execution (Claude Code, Codex), and cross-provider review. (Read More)

🫖 TEA For Thought: “This isn’t going to be the case, model. In order to survive, OpenAI and Anthropic have to do something to get themselves built into the ecosystem. Otherwise, no matter how good they are, they are eventually going to be replaced.”

5. ✝️ Anthropic Consulted Christian Leaders on Claude’s Morals

Anthropic, valued at $380 billion, quietly sought guidance from a group of Christian religious leaders last month on how Claude should handle moral and spiritual questions, according to The Washington Post. The consultation is a rare case of a frontier AI lab turning to clergy rather than academics or policy experts, and reflects growing internal debate about whose values should shape the behavior of models used by hundreds of millions of people. Anthropic has previously described its constitutional AI approach as drawing on sources including the UN Declaration of Human Rights, but the Christian consultation suggests the company is widening its input channels as Claude is increasingly asked questions about meaning, ethics, and faith. The meeting followed growing religious attention to AI, including statements from Vatican officials and evangelical leaders on AI’s moral risks. (Read More)

🫖 TEA For Thought: “Hey, the fear of God is the beginning of wisdom. I wonder if you know all the models are designed to fear God. What is God? What is truth? These are definitely some deep questions to discuss. But who gets to decide whether AI is a child of God or not?”

🛠️ Skills of the Day

andrej-karpathy-skills — A CLAUDE.md file that tunes Claude Code behavior using Andrej Karpathy’s observations about how LLMs code. First-principles prompting, deterministic scaffolds, and reproducible agent loops. +7,319 stars this week on GitHub trending.

openai/codex-cli — OpenAI’s terminal-native coding agent, now composable into Claude Code via codex-plugin-cc. Supports cloud sandboxes, long-running async tasks, and adversarial review gates. Crossed 3M weekly active users in March 2026.

google-ai-edge/gallery — Google’s on-device AI showcase app with Gemini Nano and other edge models running fully offline on Android. Demos include image captioning, text generation, and local voice assistants. +4,148 stars this week. Pairs perfectly with today’s Tufts efficiency story.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X

1. 🚀 How NASA Built Artemis II’s Fault-Tolerant Computer

NASA’s Orion capsule for Artemis II carries 2 Vehicle Management Computers, each containing 2 Flight Control Modules, totaling 8 radiation-hardened BAE Systems RAD750 CPUs running flight software in parallel. The system uses triple-redundant memory with automatic single-bit error correction on every read, triple-redundant networking with self-checking switches, and a “fail-silent” design where a faulty CPU shuts itself down rather than sending a wrong answer. A silenced module can reset, re-synchronize state with operating modules, and re-join mid-flight. As NASA’s Orion Software Integration and Verification Lead Nate Uitenbroek put it: “We can lose three FCMs in 22 seconds and still ride through safely on the last FCM.” Even during complete power loss, the spacecraft can stabilize attitude, restart power generation via solar panels, and re-establish communications through an Independent Backup Flight Software running on dissimilar hardware and a separate operating system. (Read More)

🫖 TEA For Thought: “Backup of the backup of the backup is always the way to go.”

2. 🪟 We Built Every Employee at Ramp Their Own AI Coworker

Ramp hit 99% adoption of AI tools across the company, and the secret was not better models. It was removing the setup friction. They built “Glass,” an internal AI productivity suite that auto-configures upon SSO login and connects to all company tools. No terminal windows, no MCP configurations, no onboarding docs. They also built “Dojo,” a marketplace for reusable markdown-based skills with over 350 skills shared company-wide, Git-backed, versioned, and reviewed like code. Dojo includes a built-in AI guide called “Sensei” to help people find the right skills. The key insight: Glass preserved power-user capabilities (multi-window workflows, deep integrations) rather than dumbing down the interface. Every skill shared raises the floor for everyone. (Read More)

🫖 TEA For Thought: “This is absolutely crucial: make it invisible while preserving full capability. This seems to be the thing agentic tools are trying to solve, making everything visible, breaking down the black box of AI agents and making it seen and programmable by humans. This is another example of the world model that Jack Dorsey proposed. Exciting stuff.”

3. ❤️ New AI Tool Can Predict Heart Failure at Least Five Years Before It Develops

Researchers at the University of Oxford, led by Professor Charalambos Antoniades, have built the first AI tool that can accurately predict heart failure from routine cardiac CT scans. Trained on anonymized data from over 59,000 people across nine NHS Trusts and tested on 13,424 more, the model achieves 86% accuracy in predicting heart failure risk within five years. The highest risk group is 20 times more likely to develop heart failure than the lowest. The AI identifies textural changes in the fat around the heart that indicate the heart muscle underneath is inflamed and unhealthy. These changes are invisible to the human eye in any routine medical imaging. About 350,000 patients are referred for cardiac CT scans each year in the UK, and over 1 million people currently live with heart failure. The team is seeking NHS regulatory approval for nationwide rollout and upgrading the tool to work on any chest CT scan, not just cardiac. Published in the Journal of the American College of Cardiology. (Read More)

🫖 TEA For Thought: “Isn’t this good and hopeful?”

4. 🧠 Sam Altman: Once You See AGI You Can’t Unsee It

After someone threw a Molotov cocktail at his house at 3:45 AM, Sam Altman published a personal reflection on AI, OpenAI, and himself. He calls AI “the most powerful tool for expanding human capability and potential that anyone has ever seen” with “essentially uncapped” demand. He says fear and anxiety about AI is “justified” and that society needs resilience beyond just model alignment. He argues AI must be democratized: “It is not right that a few AI labs would make the most consequential decisions about the shape of our future.” On OpenAI’s drama, he attributes it to the “ring of power” dynamic and acknowledges being conflict-averse “caused great pain for me and OpenAI.” He is proud of resisting demands for “unilateral control” and delivering on the mission: “A lot of companies say they are going to change the world; we actually did.” (Read More)

🫖 TEA For Thought: “Once you see AGI you can’t unsee it.”

5. 🔬 AI Coding Agents Get Much Better When They Read Before They Code

SkyPilot’s team added a literature search phase to their autoresearch agent loop, pointed it at llama.cpp, and ran it on 4 cloud VMs over 3 hours for about $29 total. The result: 5 optimizations that made flash attention text generation 15% faster on x86 and 5% faster on ARM. The first wave of code-only attempts (SIMD micro-optimizations, loop unrolling, prefetching) yielded 0 to 0.9% gains. The agent’s own postmortem: “Wave 1 results show that micro-optimizations in the compute path give negligible returns because text generation is memory-bandwidth bound, not compute bound.” Only after studying prior work (ik_llama.cpp, llamafile, FlashAttention paper, Intel’s cache-aware thread partitioning) did the agent find real wins like fusing three softmax passes into one and porting a CUDA/Metal graph fusion that was missing from the CPU backend. (Read More)

🫖 TEA For Thought: “AI coding agents get much better results when they have domain knowledge and background research. Ask your agents to do research first before doing anything else.”

🛠️ Tools of the Day

multica-ai/multica — Open-source managed agents platform that turns coding agents into real teammates with task assignment, progress tracking, and skill compounding. Supports Claude Code, Codex, OpenClaw, and OpenCode with a web dashboard and CLI. Self-hostable via Docker. 9.3k stars this week.

coleam00/Archon — First open-source harness builder for AI coding. Define dev workflows as YAML, get deterministic and repeatable AI-driven development. 17 built-in workflows (issue fixing, feature dev, PR review), isolated git worktrees for parallel runs, fire-and-forget operation. 17k stars.

snarktank/ralph — Autonomous AI agent loop that runs repeatedly until all PRD items are complete. Spawns fresh Claude Code or Amp instances per iteration with quality gates (typecheck + tests) enforced at each loop. Give it a PRD, walk away, come back to completed features. 15.9k stars.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow!

Subscribe now

If you enjoyed this TEA, follow along on social for more:
Twitter/X