Daily TEA – Coinbase Hands Your Crypto Wallet to an AI Agent

Coinbase launches agents that trade and pay for their own research over x402, Microsoft quietly bars its own staff from Claude Fable 5, Ramp starts renting out the AI finance engineers it built for it

Jun 12, 2026

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

1. 🪙 Coinbase Hands Your Crypto Wallet to an AI Agent That Trades and Buys Its Own Research

Coinbase launched its own trading agents on Thursday, just days after Robinhood rolled out agents that trade for users. The agent connects to your main Coinbase account, or to a separate sandbox if you would rather not give it the keys, and can rebalance a portfolio, follow an investment thesis you hand it, or weigh in on a one-off trade. It works through Coinbase Advanced, the pro platform with TradingView charts, and can trade crypto spot markets and derivatives now, with equities and prediction markets planned next. The twist is payments: the agent runs on x402, the open payment protocol Coinbase launched last year with AWS, Anthropic, Circle, and Near, so it can pay for premium research APIs and on-demand compute with no login or subscription. It also runs inside ChatGPT or Claude through a Coinbase MCP server. Custom guardrails like maximum trade size and spending caps are coming, and the Financial Stability Board this week called for stronger safeguards on agentic AI in finance. (Read More)

🫖 TEA For Thought: “The day has come.”

2. 🔒 Microsoft Bars Its Own Employees From Anthropic’s Claude Fable 5 While It Ships the Same Model to Customers

Microsoft has restricted its own employees from using Anthropic’s newly launched Claude Fable 5 while its legal and compliance teams review how the model handles data. The sticking point is Anthropic’s retention policy for the Mythos class of models, which Fable 5 belongs to: prompts and outputs are kept for 30 days for safety monitoring, and anything flagged by Anthropic’s safety systems can sit on its servers for up to two years. That clashes with the enterprise data rules Microsoft has in place, and the company is worried staff could feed customer data, corporate details, or other confidential information into the model. Fable 5 is absent from the model picker for Microsoft’s internal GitHub Copilot, and workers have been told to lean on Microsoft’s own tools instead. The irony is that Microsoft rolled the very same model out to its GitHub Copilot and Foundry customers without hesitation. It also follows Microsoft moving its developers off internal Claude Code licenses toward GitHub Copilot. (Read More)

🫖 TEA For Thought: “It is hard to do the right thing, and sometimes when you push too far the result is not what you intended. So what is Anthropic actually trying to achieve here? Safety and security by controlling and censoring everything? And if so, who gets to audit Anthropic?”

3. 🧮 Ramp Starts Renting Out the AI Finance Engineers It Built for Itself

Ramp introduced Applied AI Solutions, a service that embeds Ramp engineers inside a customer’s finance team to build and ship AI workflows on the Ramp platform in a matter of weeks. The pitch grows straight out of Ramp’s own operation: its finance team leans heavily on production AI agents running on a “Finance Intelligence” layer that maps GL accounts, policies, and reporting logic to how the business actually works, letting Ramp run with a fraction of the headcount its scale would normally require. The new team runs a structured discovery process, connects data wherever it lives (ERPs, warehouses, cloud storage, even paper workflows), builds that same intelligence layer for the client, and deploys agents tied to the customer’s KPIs. Ramp says it is model-agnostic, routing each workflow to whichever model performs best on cost and accuracy across the 70,000-plus businesses and 200 billion dollars it processes annually. The backdrop: across Ramp’s customers, AI token spend is up 13x since January 2025, yet while 87% of CFOs call AI critical, only 21% report measurable results. (Read More)

🫖 TEA For Thought: “This is a smart business model. You build the team that implements the tech for yourself first, then you sell that exact capability to everyone else. Ramp is essentially packaging its own finance-specialized forward-deployed engineers and selling them as the product.”

4. 🍔 DoorDash’s New ‘Ask DoorDash’ Lets You Order Dinner From a Prompt or a Photo

DoorDash launched an AI chatbot called Ask DoorDash that lets people order food and groceries using plain-language prompts and photos instead of scrolling through restaurants and stores. You can tell it what you are in the mood for, drop a recipe link to pull the ingredients, or snap a photo of a cookbook page or grocery list, and it will assemble the cart with the right items and quantities, even prompting you to check whether you already have staples like sugar and butter. For restaurants, a query like “a filling dinner for a family of four” surfaces options with a short blurb on why each fits, and you can narrow it further to “kid-friendly vegetarian spots with mild options.” It also handles reservations, taking a request like “a table for two downtown for date night around 8 PM” and surfacing places with availability. The chatbot is rolling out on iOS in select regions and will reach more of the US in the coming weeks, joining similar moves from Uber Eats and Instacart. (Read More)

🫖 TEA For Thought: “A pretty good and authentic pivot. The only real question is whether the cost of image generation, LLM calls, and the rest actually pencils out.”

5. 🍎 Apple Shelves the New Siri for All of Europe With No Timeline to Bring It Back

Apple confirmed it will not ship its overhauled, AI-powered Siri as part of iOS 27 and iPadOS 27 in the European Union, blaming conflicts with the Digital Markets Act, and it gave no timeline for when, or whether, that changes. According to Apple, the DMA would force it to give any AI system “nearly unlimited access to a user’s device” plus the ability to act autonomously “without a user’s ongoing visibility and control.” Apple says it proposed a “Trusted System Agent” intermediary and an 18-month rollout plan, but the European Commission rejected its proposals. Craig Federighi said Apple hopes to bring Siri AI to the EU eventually but that the Commission’s “refusal to engage constructively” leaves it without a timeline. EU developers also cannot test the Siri AI features on the betas, and the new Siri and Apple Intelligence features will not ship in China either, citing regulatory work. EU users can still get the revamped Siri on macOS 27, visionOS 27, and watchOS 27. (Read More)

🫖 TEA For Thought: “The decline of Europe in a single headline: rare innovation, extreme regulation.”

🛠️ Skill of the Day

The Agent Guardrail: before you hand an AI agent your wallet, your inbox, or your production keys, make it design its own safety rails first, so you find the failure modes before they find you (ask Coinbase and the Financial Stability Board why this is suddenly the conversation).

You are a risk engineer, not my assistant. I am about to let an AI agent act on my behalf. Before it touches anything real, your job is to design the guardrails that keep a bad day from becoming a disaster.

THE AGENT'S JOB: [DESCRIBE WHAT THE AGENT WILL DO, E.G. "REBALANCE MY PORTFOLIO," "ANSWER CUSTOMER EMAILS," "DEPLOY CODE"]
WHAT IT CAN TOUCH: [ACCOUNTS, MONEY, DATA, SYSTEMS, PEOPLE]
WORST REALISTIC LOSS: [MONEY, DATA, REPUTATION, TIME]

Do this in order:
1. List the 5 ways this agent could cause real damage, ranked by how badly it would hurt, not how likely it is.
2. For each one, give me a concrete hard limit that would cap the damage (a spending ceiling, a read-only scope, a human approval step, a rate limit, a kill switch).
3. Tell me which single action this agent should NEVER take without a human pressing the button, and why.
4. Design the "panic test": one thing I can check each day in under a minute to know the agent has not gone off the rails.
5. Write the agent's operating rules as a short instruction block I can paste into its system prompt, written so a careful agent would refuse to cross the lines you set.

Assume the agent is competent but literal, and that anything you do not explicitly forbid, it will eventually try.

Paste into ChatGPT, Claude, or your tool of choice, and replace the bracketed bits. Run it before you give any agent permissions, not after.