When AI Writes the Code, Who Cleans Up?

AI coding agents can write software faster than teams can review it. This talk gives you a practical blueprint for keeping an agent-built codebase coherent: translating architectural preferences into rules agents can evaluate, running recurring repository scans, and turning the highest-value findings into focused pull requests without creating another backlog for developers to babysit.

I’ll show how we built Codex GC, an internal maintenance system that encodes “golden principles” about how a codebase should work. Its agents look beyond passing tests for duplicated utilities, inconsistent abstractions, outdated conventions, and architectural drift. They rank findings by confidence and impact, then propose remediations that humans can review and merge. I will also speak to fully closing the automated loop and how this system interacts with AI PR reviews.

This talk will also dive into the complexities of teaching the system when to act, when to abstain, and how to learn from rejection, allowing the system to accrue team-specific tacit knowledge over time and compound.