AI Builders Digest

Coinbase is opening bank accounts for AI agents because cards can't do 1-cent payments

About 76% of the agent ecommerce transactions Coinbase sees are under 30 cents, which is the flat floor on a card transaction, so the economics simply don't exist on traditional rails. Coinbase now hands any agent a self-custodial wallet from a single pasted prompt, no KYC, and is pushing the x402 protocol it donated to the Linux Foundation with Google, Cloudflare and AWS on board. "We don't want the AIs to be unbanked... if we want to bank the AIs, they deserve financial services as well." Internally the bet is a "brain" per team and per repo, fed every incident, financial control and A/B test, and the discipline is that when a human corrects an agent's PR that correction goes back into the brain rather than just into the diff. Also notable: 88% of Coinbase revenue is now non-Bitcoin trading, prediction markets hit a $100M revenue run rate within months of launch, and New Limit's first phase one trial, in alcoholic liver disease, starts next year.

• #agents
• #payments
• #products

The personal agent race scorecard: Muse to win, ChatGPT losing on UX split

Muse is poised to take the lead because the app is intuitive and Meta is pushing it across every property, and it should end up bigger than Threads. ChatGPT still holds the crown on raw users and arguably beats Muse feature by feature on models, computer use and voice, but one product cannot serve both work and personal use, and the Work/Codex split is confusing enough to drive people elsewhere. Google is the dark horse since every agent lives off Gmail and Calendar data, yet Spark is buried as a secondary tab in Gemini. The real unsolved problem is multiplayer: you cannot add a spouse to a Muse chat or loop a coworker into a thread, and whoever cracks that gets the next big unlock. Practical advice: design your personal skills and files to be portable across harnesses, or you will spend half your time migrating between tools.

• #agents
• #products

An agent built an ephemeral deployment to chase one mobile rendering bug

Given a vague complaint that something looked wrong in an in-app mobile browser, the agent went all the way down: reproduce, simulate, fix, deploy, verify, including spinning up an ephemeral Vercel deployment to throw at an iPhone simulator. Humans never do this. We run out of energy, we extrapolate, we hope it works out. The claim is that software quality and performance are about to hit levels we have never experienced, purely because QA no longer has an attention budget.

• #agents
• #products

Jev is the whole conversation now, and Instinct was the whole conversation last week

A dry note on how fast AI attention cycles turn over: everyone is obsessed with Jev now, but people could not shut up about Instinct as recently as the middle of last week. Simpler times.

• #industry

OpenClaw can now FaceTime you

The claw gained a FaceTime capability, alongside a new benchmark drop. Small feature, notable direction: agents reaching out over a live video call rather than waiting in a chat window.

• #agents
• #products