Your Agent's requirements.txt Is a Lie

Coding agents write requirements.txt with total confidence and no memory of what is actually installed: inventing package names, guessing version pins, and blending APIs from different library eras. CI either fails opaquely or silently resolves to whatever the cache happens to have lying around. This talk treats that as a caching and provenance problem. Drawing on build-acceleration work with distributed compile farms, I will show what it takes to pin, content-address, and verify Python environments an agent proposes before any of that code reaches a real sandbox: hash-locked dependency resolution that treats the agent's guess as untrusted input to the cache key, and a fast-reject path that catches hallucinated package names before a slow install ever starts.