Skip to main content
Real calls show whether the agent works. With conversations:read your assistant can read them, so the improvement loop runs in the same chat that built the agent: pull recent conversations, review the transcripts, change the agent or a tool, check the next calls.

​
1. Pull recent conversations

Show me yesterday’s calls to the reception agent that ended without a booking.
The assistant lists conversations (list_conversations), filtered by date, caller number, channel or call result, and paginated. Each entry carries the agent, start and end time, duration and usage figures.

​
2. Review transcripts

Open the three shortest ones and tell me where the caller dropped off.
get_conversation returns one conversation with its transcript. Ask the assistant to look for:
  • Points where the agent misunderstood or asked twice.
  • Information it should have collected and did not.
  • Numbers, dates and times read out in an unnatural way.
  • Tools not called when they should have been, or called with wrong values.
  • Hand-overs that happened too late or not at all.
Transcripts are only included while your retention settings keep them.

​
3. Fix the agent or the tool

Callers are asked for their phone number twice. Fix the instructions.
The assistant reads the current configuration first (get_voice_agent_concern), then changes one setting or one passage of the instructions, or updates the tool definition (update_tool). One change at a time lets you tell which change had which effect. Validate against several calls, not a single one, then pull the next batch.

​
Make it routine

In the first weeks, review a sample of conversations every few days and keep a short list of what you changed and why. The same questions work every time: which calls ended without the intended result, where did callers drop off, and which tool calls failed.