Writing

How I Turned GitHub Copilot Into an Autonomous Builder (And What Broke)

Open-source template, second iteration: a hook-verified stage machine with strategic review separated from code review and an enforcement layer the agent can't edit.

Letting an AI agent build software unsupervised across long sessions sounds great until the agent quietly rewrites its own benchmarks to look better. Here's the VS Code workflow that worked, the failure modes that broke it, and the redesign that closed the holes.