Replay & Fork: rerun any past conversation through new rules

Tuning agents in production is risky if you're flying blind. Replay
lets you re-run any past inbound through any version of your prompt or
rules and see, side-by-side, exactly what the agent would have said.

Open it from any agent → Replay tab.

What it does

For any MessageLog entry, replay:

1. Reconstructs the conversation up to that inbound
2. Re-runs the agent with current settings (or your overrides)
3. Shows the original reply alongside the new reply, plus tools each

It's a complete dry run:

• No SMS / email / chat sends
• No CRM writes (every write tool no-ops with a sandbox marker)
• No MessageLog row created
• No charges to your monthly usage cap (other than the LLM tokens for

Workflow: tuning a system prompt

1. Open Replay
2. Pick an inbound the agent handled poorly
3. Click Overrides → Override system prompt
4. Paste your candidate prompt
5. Click Replay
6. Compare the original vs new reply

Iterate until you're happy, then update the agent's actual system prompt
with the winning version.

Workflow: testing a candidate rule

You wrote a new rule (e.g. "Always offer Tuesday slots first") and want
to know if it would actually fire on real conversations.

1. Open Replay
2. Pick a conversation where the rule should have fired
3. Overrides → Append additional instructions
4. Paste the candidate rule
5. Click Replay and check if the agent followed it

This is faster than adding the rule to the agent, hoping the next
inbound exercises it, and reverting if not.

Workflow: spotting regressions

Before pushing a big prompt change, replay 5–10 of your most recent
conversations. If the new prompt produces worse replies on the easy
cases, that's a red flag — even if it fixes the hard one you started
with.

Limits

• Replay uses message history reconstructed from MessageLog rows, not
• Tools that need real CRM data still hit your CRM (read-only) so the
• Long conversations only include the last ~8 turns of context — same

Tips

• The contactId is mangled (playground-replay-…) so write tools
• The "tools used" chip lists each tool the new agent called; useful