You cannot govern an AI agent against meaning that was never formally defined.

Every business application defines meaning privately — reconciled N×M times, heuristically, at the edges. AI agents turn that from an inconvenience into a structural failure.

A few months ago, I wrote about adapting the Karpathy loop for production prompt optimization. Since then, the biggest lesson has been that the hard part is not the prompt. It is the measurement.

The updated system moves the metric closer to the real user promise and resolves the hold-out problem by making verification part of corpus growth itself.

The result is a loop that is more honest, more robust, and much harder to fool.

In early 2025, Andrej Karpathy described a simple but powerful pattern for autonomous software improvement he called "autoresearch." The core idea: give an LLM agent a frozen evaluation harness, a bounded editable artifact, a natural-language specification of what "better" means, and a ratchet that commits on improvement and reverts on regression. Run it overnight. Collect improvements.