How do you check that your coding agent's "done" is actually done?
retained text
I've been using Claude Code (and trying Codex) on real projects, and the thing that bites me most isn't bad code. It's confident "all tests pass, fixed!" messages that turn out to be wrong. The patterns I keep seeing: • It ran the tests, then made **one more edit** and reported the old result • A test got **skipped or loosened** instead of the code being fixed • The fix only works for the **exact input** the test uses What I do now: ask it to re-run the full suite after its last change and paste the summary line, and glance at git diff --stat for any test files it touched. Curious what others do. Do you trust the agent's summary, run tests yourself, use hooks, or something else? Have you ever merged something an agent said was done that wasn't?
document fetched 2026-11-16 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xx0mnh/how_do_you_check_that_your_coding_agents_done_is