Metrics & Monitoring
ANTS Platform computes performance metrics from your traces. Every trace and observation has a start and end time, and every observation has a level, so latency and errors can be broken down by agent, step, model, user and environment.
The views below are under Gauge AI and need Gauge AI read access (Owner, Department Admin, Security & Compliance and Cost Analyst).
Latency
| Where | What it shows |
|---|---|
| Gauge AI → Latency Dashboard | p95 latency by trace name, by model and by observation level; p95 time to first token by model; average output tokens per second by model; max latency per user. See Cost & Latency Dashboards. |
| Gauge AI → Latency & Analytics | p50, p90, p95 and p99 tables for traces, generations and spans by name, and model latencies over time from p50 to p99. See Cost & Performance Metrics. |
Latency is the duration of a trace or observation. Time to first token needs the completion start time on streamed generations, which the SDK integrations record.
Errors
Observations carry a level: DEBUG, DEFAULT, WARNING or ERROR, with an optional status message.
- In Discover AI → Tracing, filter by Level to list traces that contain errors.
- On the Latency Dashboard, P 95 Latency by Level (Observations) shows whether failing calls are also the slow ones.
- For an error share over time, clone a dashboard and add a widget that counts observations grouped by level. See Custom Dashboards.
There is no built-in error rate, success rate or throughput metric; derive them from counts as above.
Measures available for custom widgets and the API
| Data view | Measures |
|---|---|
| Traces | count, observations count, scores count, latency, total tokens, total cost |
| Observations | count, latency, streaming latency, input tokens, output tokens, total tokens, output tokens per second, tokens per second, input cost, output cost, total cost, time to first token |
| Scores (numeric and categorical) | count, value (numeric) |
Aggregations are sum, avg, count, max, min, p50, p75, p90, p95, p99 and histogram. The same query model is available through the public Metrics API, GET /api/public/metrics. See the API Reference.
Next steps
- Alerting & Notifications: what can notify your team.
- Performance Optimization: act on slow agents and models.