Daily TEA – Street Smarts Are Coming, Agent Swarms Prove Their Work, AI Divide Is Here
Science agents still flunk real experiments, 20 parallel agents beat human experts, Chrome copies Perplexity, hybrid workflows win, and 74% of AI gains go to 20% of companies
Hello, dear TEA-mates! Here is what you need to know today.
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.
If you enjoyed this TEA, follow along on social for more:
Twitter/X