How to verify what an AI agent actually changed: Reddit's field-tested checklist

falafel777 · reddit · 2026-10-23

A Reddit roundup of answers to 'how do you verify what an agent really did before merging':

Confine writes to a container or worktree; out-of-repo writes should fail loudly
Before/after snapshots miss the mutate-and-restore case: agent edits outside config, runs tests, reverts—both ends look identical, the green was never real
Path filters miss npm config set and git config --global (no path in command); symlinks fake in-repo paths
Test config counts as test changes: retries 0→2, added .skip, raised timeout all keep suites green; if CI the agent didn't trigger disagrees, it didn't pass
Read back after every write, not at the end; never blindly retry timed-out writes (duplicates)

Key line: a check you skip when busy isn't a check. Author still seeks sandbox-free ways to catch mutate-and-restore.

More from coding & agent