Replay voice calls

This walks through replaying a single production voice call: finding it, seeing what the replay carries over from the original, and reading what the run hands back. For what replay actually reconstructs and why that matters, see Replay.

Find the call

Open the call from Explore sessions & users and start a replay from there, as a trace since one voice call is one trace. If you’re after a whole multi-call conversation instead of a single one, that’s a session replay, and it works the same way from that point on.

What carries over

Starting the replay doesn’t touch the original call. It creates a copy for you to work against: a voice agent definition carrying the provider, model, and assistant configuration that call actually ran on, and a scenario built from its transcript. Edit either freely. Nothing you change here reaches the agent that’s still taking real calls.

Run it

The agent and scenario land in an ordinary simulation run, and a voice run starts placing calls the moment it’s created, same as any other. Run a voice simulation covers that wizard in full.

To test a fix, edit the recreated agent definition, or point the run at a newer version, then place a fresh call rather than only rescoring the old one: open the run and rerun it with Run test + Evals.

What comes back

Once the new call finishes, open it and click Compare with baseline to see it next to the original. You get both transcripts side by side, both recordings to play back, and the call metrics for each: duration, agent latency, talk ratio, words per minute, and interruption counts. That comparison is what tells you whether the change actually moved anything, not just that the new call completed.

Dive deeper