[PROPOSAL] Cookbook: An agent loop that improves its own tooling between runs

Most agent examples run a task and stop. A more interesting loop is one where the agent improves its own tooling between runs: it does the task, notices where it was slow or where it had to redo work, edits the small script or prompt it used, and the next run is faster or cleaner because of that edit. Over several cycles the scaffolding the agent built for itself starts doing the heavy lifting.

I would write this up as a concrete, measurable loop rather than a philosophy. The point of the cookbook is to show that the improvement is real and to measure it, not just to assert that self-improvement is possible.

What the notebook would cover:
- A fixed repeatable task with a clear score (time, steps, or a correctness check against a reference).
- The loop: run, have the model critique its own run for the single biggest friction, apply one change to its tool or prompt, run again. One change per cycle so the effect of each is attributable.
- The measured curve: the score across N cycles, including the cycles where a change made things worse and was reverted, so the picture is honest rather than a clean line.
- The two failure modes to call out: changes that overfit the one example, and a loop that keeps editing without converging. Simple guards for each (a held-out second task, and a stop rule when a few cycles in a row find nothing).

It runs on the messages API with a small task harness. I would target the agents section. Happy to anchor it to a specific task (a parser, a small data-cleaning job) if a concrete example reads better than a general harness.