1. Open the agent
From Agents, open the agent. The detail page shows a conversation-volume trend chart and an hour-by-day heatmap, so you can tell whether today’s dip is a blip or a trend.2. Drill into the conversations
The detail page lists the agent’s most recent conversations. Open a conversation to read the transcript next to the system prompt and listen to the recording. To narrow in, filter the conversations by ended reason, has-audio, or date range. The agent’s system prompt (and its flow) are on the detail tabs too, so you can check its behavior against what it was actually told to do.3. Work the findings
When open findings for an agent pile up, the agent surfaces a prompt to review them. From there, triage the findings and capture a change for the fix.Audit programmatically
- MCP —
get_agent,list_conversations, andget_transcriptfor reviewing the agent’s conversations and transcripts; aggregate the returned conversation records in your client when needed (e.g. failure counts by ended reason).
Related
- Agents — the concepts behind an audit.
- Conversations — the conversations you’re drilling into.
- Findings — the recurring issues you promote and fix.

