iterate-agent-harness — Failure → Durable Guard

Degree of freedom: MIXED. Failure-class and smallest fix [HIGH freedom]; guard-fails-before-fix and the report shape [LOW freedom — run exactly].

When an agent run ends wrong, the fix is not just to redo the task — it is to change the harness so the whole class of failure cannot recur silently. This skill converts a single failure into a durable guard plus a regression check.

A failure you cannot reproduce, you cannot prove you fixed. Capture the failure mode, add a guard that fails on it, then make the guard pass with the smallest durable change.

How to reason

Observe — happened vs should-have; transcript, state file, diff, CI
Interpret — routing miss, missing guard, or a gate that passed wrongly
Classify — premature stop / false completion / reward hacking / scope / handoff / context loss
Guard — add a guard that fails on the unfixed behavior, then the smallest durable fix