The loop learned slowly from corrections and faster from verdicts.where a self-improvement loop's learning signal comes from, measured on one governed agent fleet through week 41, and what to feed your own.

Every agent program that learns from its own operation is already choosing what the loop consumes: the judgments people type when an agent goes wrong, or the machine's own verdicts about execution, a timeout, a non-zero exit, a refused command. I ran both channels on one governed fleet and measured the yield of each through calendar week 41. Adjudicated corrections produced one new behavioural rule from about twenty candidates escalated over eleven weeks and did real work as a filter; the July A/B this series owed its readers measured the mined generic rules as neutral to harmful, and three were archived. Execution friction, ranked by how many sessions paid for each failure and validated on held-out weeks, found the fleet's most expensive recurring defect and re-found it unprompted on a second corpus. The loop's first unattended selection was rejected with reasons; its first learner-selected fix measured 28 tool calls down to 11 across six trials per arm on one task, under a rule frozen before the trial; a second candidate measured no benefit and was withheld. The five-state ledger with every column dated, what is and is not established, and a one-week exercise plus a checklist an enterprise agent program can adopt now.