One Human, Many Agents. How I Run an Entire R&D Department Alone

For the past year one human has run a software R&D department in which the four developers, working as two teams of two on separate servers, the three code reviewers, the arbiter and the QA team are all AI agents. The department has shipped hundreds of gated changes into a production engine. It was not the models that made this work. It was the rule book.

This talk is about how agents that share no memory and no trust are made to coordinate:
every message is a file with a schema, every verdict is bound to one commit, the board leads the work instead of recording it, and nobody grades their own work. Two steps are new to the industry. Reviewers are hired by measurement, on a 65-case adjudicated benchmark that exposed how flattering the earlier benchmark had been. And the AI tester runs behind four mechanical locks that make it physically unable to touch source code.

Every rule is presented with the incident that wrote it. The human’s remaining job turns out to be a PO’s job: decide scope, requirements and high level architecture, adjudicate severity and approve changes.