Part 1 of 4 · 1 min read
Overview
Coding is the agent domain with the cleanest feedback signal (tests, types, linters) and with a real risk of confident wrong changes. It teaches the pattern behind every tool-using agent: verify with something other than the model's own opinion, and bound what it can change.
About the example, and what is real
Real: the workspace (guarded read, search, exact-match edit, subprocess test runs with timeouts, diff), the review gate, the loop and the held-out check, and the 20 tests in code/mini-coding; every number below comes from actually running them. Scripted: the candidate patches are written by me, in a fixed order; there is no model, so the results say nothing about how often a real model makes each mistake. Not real: six tiny functions, not a repository; the test runner is Python unittest in a temporary directory, not a hardened sandbox.