Same prompt, Monday vs. Tuesday — different answer. The agent rewrites the wrong file. It calls a function that was deleted three months ago. We blame the model, but most of the time the problem is upstream: the context the model sees is non-deterministic. Whatever files it stumbles on, in whatever order, with whatever conventions it has to guess at — that's what shapes the output.

This talk is about engineering context as a first-class concern. Why "let the agent grep around" produces unstable results. What determinism actually means when the input is a token budget. And five principles that turn a flashlight in a dark room into a briefing packet for any LLM.