Why do tool outputs need to be captured instead of called again live?

Tool outputs need to be captured, not re-fetched, because the systems behind them keep moving. A database row updates, an inventory count changes, a search index reindexes, so calling the same tool with the same arguments an hour later can return a different result than the one the agent actually saw. If replay re-calls tools live, you’re not replaying the failure anymore, you’re running a new session against current state that happens to share a starting point. The fix you’re testing might pass only because the data changed underneath it, not because the fix works. Recording each tool’s output at the time it was returned, and replaying it verbatim instead of the live call, is what keeps the replayed case identical to the one that actually failed.