Observability
Monitoring
Dashboards for the numbers that predict a bad customer experience — answer rate, response latency, error rate, and tool health — plus alerts so you hear about problems before your customers do.
The four signals
Latency
End of caller speech to first agent audio, p50 and p95, broken down by pipeline stage.
Availability
Answer rate for inbound and connect rate for outbound, per number and per agent.
Error rate
Share of calls ending in a pipeline, tool, or carrier error.
Tool health
Success rate and p95 duration for every configured tool.
Healthy baselines
| Metric | Healthy | Investigate | Incident |
|---|---|---|---|
| Response latency p95 | < 1.0s | 1.0–1.8s | > 1.8s |
| Inbound answer rate | > 99% | 97–99% | < 97% |
| Pipeline error rate | < 0.5% | 0.5–2% | > 2% |
| Tool success rate | > 98% | 95–98% | < 95% |
| Webhook delivery rate | > 99.5% | 98–99.5% | < 98% |
Latency breakdown
The latency panel splits each turn into transcriber endpointing, model first token, tool wait, and TTS first byte. Optimize the largest bar — teams routinely rewrite prompts when the real cost is a 900ms CRM lookup.
Alerts
- 1
Open Data → Monitoring → Alerts
Alerts are workspace-scoped and can be filtered to a single agent or number.
- 2
Pick a metric and threshold
Choose a window long enough to avoid noise — five minutes is a good default.
- 3
Choose channels
Email, in-app notification, Slack, or a webhook to your on-call system.
- 4
Set severity
Warning routes to notifications; critical pages your on-call channel.
{
"alert": {
"name": "Agent latency degraded",
"metric": "response_latency_p95",
"condition": "> 1800",
"window_minutes": 5,
"scope": { "agent_id": "agt_123" },
"severity": "critical",
"channels": ["slack:#oncall", "webhook:pagerduty"]
}
}Recommended starter alerts
- Error rate > 2% for 5 minutes — Catches model and carrier outages before they cost you a morning of calls.
- Answer rate < 97% for 10 minutes — Usually a number or trunk problem, not an agent problem.
- Any tool below 95% success — Your calendar or CRM is failing and the agent is improvising.
- Webhook delivery failures — Your own endpoint is down and downstream systems are silently missing data.
- Credit balance below one day of usage — Prevents insufficient-credit endings mid-campaign.
Status and uptime
Platform-wide incidents are published on the Katexs status page, linked from the docs header. Workspace-level dashboards show your own uptime separately, so you can tell a platform incident from a misconfiguration on your side.