Corey James- 3 weeksor less for the native C++ port of ActiveKV under a seven-criterion definition of done
- 24performance promotions through the procedure in four days
- 60+locked decisions in the Agent Swarm master plan before parallel work
Most of the code behind the systems on this site was written by AI coding agents. I lead that work as an architect: I keep the agents that build apart from the agents that judge, and make the calls when the evidence is disputed. I write the definition of done, the intake standard and the promotion procedure before work starts, and each new ruling becomes standing doctrine. The method is the story Directing AI coding agents at program scale.
Taking the memory out of the model
AgentOS started from a diagnosis. Long agent runs drift because the lead agent has to hold the procedure, the plan, every teammate's state and every piece of evidence in its context, and past some horizon it drops one. I moved that memory out of the model.
About a hundred hook modules turn policy into refusals. Builders and judges come from different model families, so one family's blind spots meet another's.
An agent's report that it is done is a claim; the system closes work on evidence a separate seat has checked.
Agent teams built AgentOS while working under it. I then ran it on real deliveries: a fourteen-day platform build the system closed only when every close item had a current passing verdict, and a port of a large open-source agent gateway into a 25-crate Rust workspace, measured against a pinned census of the original source. See Why long agent runs fail and Separation of duties for AI work.
Locked decisions and gatekeepers
Agent Swarm is where I first worked the method out. I wrote a master plan of locked decisions, froze interfaces before parallel work began, and used gatekeeper agents that had never implemented what they judged.
When agents drift, I change the machinery so the drift is impossible.
A port judged bit for bit
ActiveKV moved from Python to a native C++ engine under a definition of done I wrote. Served output had to be bit-identical wherever the reference was deterministic, and every line of Python left behind needed a framework contract to justify it. During the port I voided over a thousand passing trials when I found a component had been registered after the code path was chosen.
Performance work then ran through a promotion procedure. See Rewriting an engine extension without changing a token.
A model built under one work order
Capabilities model was one work order with an independent quality seat, covering data, training, a custom decoder, serving and a frozen evaluation. See The baseline that tied us.
Held open until the evidence closed it
FAFO Core and Function Forge includes an API integration platform built in gated phases with a conformance suite. When the agents reported it finished, I held it open for independent close evidence. See Closed by evidence.
FAFOHive gained three social-platform integrations in about two days, each built against a written standard every tool had to meet before release.