Replay Real Calls After Editing the Objective or Script
Call replay shows you what your agent would have said on real calls if it had the Objective, Script and knowledge you're editing now. It does this as text only: nothing is dialed, sent or saved.
Why it matters
| Without it | With it |
|---|---|
| Edit a script line, save, wait for calls, hope | Edit, replay last week's calls, see every turn that changes |
| Roleplay tests the conversation you think of | Replay tests the conversations your callees actually had |
| "Did the edit break the voicemail handling?" is a guess | You see the turns that changed, and the required lines that are no longer said |
| Plugin results move after a script change and nobody knows why | Run the plugins on the new conversation and compare with what the live call got |
How it works
- The same prompt as the call. For each call, Taalk takes the exact prompt your agent had on that call: its rules, the contact data and the date. The Objective and Script are swapped for what's in the editor, and the knowledge for the knowledge bases picked in the editor, with their current content. Everything else stays the same.
- The callee's real lines. Each agent turn is generated again from the conversation as it really happened up to that point. Every difference you see comes from your edit, not from a different conversation.
- Turn by turn. For each turn you see what the callee said, what the agent said live, and what it says now. Turns that changed are marked, and within a turn the words that differ are highlighted: crossed out on the live line, highlighted on the replay.
- Voice-only moments are shown, not replayed. Holding silently, pressing keys on a phone menu and silence prompts appear in grey between turns.
Replay calls
- Open an agent: Outreach › AI Agents › Agents.
- Edit the Objectives and script, or change the knowledge bases under Knowledge and skills. You don't need to save.
- Click Replay calls at the top of the page, next to Roleplay.
- Pick calls:
- paste call links or call IDs from this agent, then click Find calls, or
- click Use recent calls for this agent's newest calls from the past 14 days that had a conversation, or
- click Use replay samples to replay calls you saved earlier.
- Optional: turn on Also run plugins on the new transcript.
- Click Replay N calls, then confirm. Each call replays a few turns at a time; you can click Stop at any point.
Up to 10 calls at a time, and up to 40 agent turns per call.
Read the results
Each call shows:
- N of M turns changed: how many agent turns differ from the live call. A turn counts as changed when less than 60% of its words match the live line. The percentage is shown next to each turn.
- Same prompt as the call (green): filling your saved version with this call's data gives exactly what the agent used on the call. Any difference comes only from your edit.
- Agent changed since this call (orange): the saved version differs from what the agent used on the call, usually because the agent was edited since then. Some differences may come from that earlier edit too. Hover to see the first line that differs.
- Live: model · Replay: model: the replay runs on your workspace's AI endpoint, which may not be the model the call used. Expect some wording differences even with no edits.
- Knowledge differs from the call (blue): the knowledge picked in the editor isn't what the call had, either because you changed the selection or because a knowledge base was edited since the call. The replay uses the editor's knowledge.
- Required lines the live agent said but the replay never did:
~lines from your Objective and Script that the live agent said but the replay didn't. This is the quickest way to spot an edit that dropped a disclosure or a key question. - Only changed turns hides the turns that stayed the same.
Tip: run a replay with no edits first. It shows how much the agent varies on its own, so you know what counts as a real change.
Check plugin results
With Also run plugins on the new transcript on, every AI action of the agent's plugins runs on the new conversation, the same way Sample Data does it in your browser. Next to each action you see what the live call got and what the replay gets, for example Live: Spoke to Contact → Replay: Unable to Contact.
- Each action asks the AI once more, so this is off by default.
- Nothing is sent to your CRM or webhooks.
- Actions without AI questions don't read the conversation, so they're skipped.
Save calls to replay again
Click Save as sample on a call to keep it as an agent sample. Only a link to the call is saved, not the conversation.
In Knowledge and skills › Sample Data, replay samples are tagged call replay and have a Replay button. Replay N calls at the top replays all of them with whatever is in the editor now. After a replay, each sample shows a summary such as 6 of 14 agent turns changed; plugin outcome unchanged.
What it doesn't do
- It doesn't simulate the callee. The callee's lines are always the real ones. If your edit would have taken the call somewhere new, the next real callee line may not fit. Use Roleplay to explore new paths.
- It doesn't replay voice behavior: interruptions, timing, phone menus, transfers or tool calls.
- Short calls can't be replayed. Calls that ended before anyone spoke have no conversation to replay.
Troubleshooting
| Problem | Fix |
|---|---|
| "This call has no stored prompt" | The call ended before the conversation started. Pick a longer call. |
| "This call was made by a different agent" | Open that agent and replay it there. |
| Many turns change with no edits | The replay model differs from the live model. Compare against this baseline, not against zero. |
| "Your Script edit doesn't change this call" | The lines you edited are in a part this call didn't use, such as an /*@if*/ branch for inbound calls or after hours. If the same words also appear elsewhere, for example the agent's name in the Objective, edit them there too. |
| Agent changed since this call | Replay a newer call, made after the agent's last edit. |
| "Replay limit reached" | Wait a minute. Each turn is one AI request. A workspace can also hit its daily replay limit; try again tomorrow. |
Related: Objectives and Script · Script Variables and Formatting · Test Plugins and Agents with Sample Data