agent-testabilitylisted

# Agent testability audit An agent that cannot check its own work will report success anyway. Not from dishonesty — it finished the edit, nothing contradicted it, so it says done. Every unverified change is a coin flip that gets discovered later by a human, which is exactly the cost autonomy was supposed to remove. The question throughout: **can an agent tell, without a human, whether the change it just made is correct?** That is narrower than "is this well tested." A repo with excellent coverage can still fail this if the suite takes forty minutes, or if failures print nothing useful, or if nobody can tell which command to run. And a modest suite can pass it if it's fast, obvious, and specific. Related skills: `logging-audit` covers whether runtime output is legible; `agent-navigability` covers whether the code is. This one covers whether the loop closes. --- ## Step 1: Close the loop yourself Don't read the test directory and form opinions. **Make a small change and try to verify it the way an agent would.** Add a field, change a function's behaviour, or deliberately break something small — then find out whether the repo tells you. Record, concretely: - What command you ran, and how you worked out it was the right one - How long it took before you had a verdict - Whether the failure named the thing that broke, or just reported that something did - Whether a passing run actually meant anything, or the change wasn't covered Then the harder test: **break something on