How Does Coinbase's Coding Agent Learn From Code Reviews?
Brian Armstrong says Coinbase's internal agent harness, Toshi, reads a markdown brain for a service, makes the code change, and sends review fixes back into the brain so the accept rate for one-shotted pull requests starts to tick up.
The brain holds the incidents on the service, the financial controls that must be enforced, every A-B test the team has run, and the full GitHub history of pull requests accepted or rejected. The files are markdown. A person asks the agent for a change. It ingests the brain and makes the change. If review finds a miss, the fix goes back into the brain instead of a one-off manual edit, so future cases improve too. Coinbase is trying to enforce the step in its software factory. Toshi can call external vendors and pay them over x402 protocols in the agentic finance economy.
Earlier the same day, he asked a high-end expensive model to plan a complex feature in three phases of 10 pieces each, then to run phase one with 10 agents in parallel and recommend cheaper models when they fit. It spun up open source agents and Grok. Two minutes before the podcast, it pinged him that all 10 pieces from phase one were done and ready for review. He used to ping the team in Slack all day. Now he tags an agent in Slack or drops the request into the harness and sends a finished pull request, which lets the CEO ship features again.
