What does replay show that a trace viewer does not?

Replay shows what would happen if you changed one thing, the prompt, a tool result, or a config setting, and re-ran the rest of the turn against the original context, so you can watch whether the model’s decision changes. A trace viewer only shows what happened: the messages, the tool calls, the outputs, laid out for reading, not what a different input would have produced. That’s the test a fix actually needs, not “does this look better” but “does the agent behave differently on the exact case that failed.” A viewer can tell you a tool returned a stale price; only replay tells you whether fixing that field would have stopped the bad response, on the same case, not a similar one.

keep reading

More on this.

Send us the traces you already emit.