Skip to main content

​
Choose what to collect

You choose what’s collected when you deploy the agent, under Monitoring. Transcripts and Traces are on by default; Recording and Logs are off.

​
Redaction

When Recording or Transcripts is on, turn on Redaction to mask sensitive data. Add a field for each kind of data to mask, with a Type (for example card_number) and the Label that replaces it (for example CARD). Choose whether to redact the transcript, the recording, or both.

​
Review sessions

Every call is a session. Open Sessions in the sidebar to see them, with each session’s duration, latency, cost, recording, and status. Select View to open a session. A session shows:
  • Overview: when it ran, total tokens, total cost, latency percentiles (P50, P90, and P95), and the system prompt used.
  • Recording: a player with seek, speed, and skip controls, when recording is on.
  • Transcript: every user and agent turn, with interruptions marked. Each turn shows its latencies (STT, LLM, TTS, end of utterance, and end to end), plus any tool calls and knowledge base retrievals. Select a turn for full details.
  • Traces: a timeline of every step the agent ran during the call.
  • Logs: the session’s log lines, with level and message.
To end a live session, select Kill Session.

​
Deployment health

Open Deployments and select your agent’s deployment, marked Agent Builder. You see its status, region, profile, and replicas, plus analytics for uptime, joining latency, and sessions over the last 5 or 7 days. To go back to an earlier deployed version quickly, select Instant Rollback. To roll back the configuration itself, use History and revert.

​
Alerts

Use Alerts to get notified when the agent’s performance crosses a threshold, for example when LLM latency goes above 3000 ms. Select Create Alert and set:
  • A name, severity (Info, Warning, or Critical), and the metric to watch.
  • The condition: an aggregation (such as P90), a comparator, and a threshold.
  • The monitoring window, and how many consecutive windows must breach before the alert fires.
  • Where to send incidents: PagerDuty, Slack, email, or a webhook.
Fired alerts appear under Incident.