Track usage, engagement, and reliability for your agent, then drill into the users, conversations, and runs behind each number.
Monitoring answers two different questions, and it helps to know which one you are asking. The overview tells you whether the agent is healthy and being used. The activity pages tell you why a specific thing happened.
Start at the overview for the trend, then follow it down into users, conversations, or logs.
The overview covers a 7, 30, or 90 day window. Each card carries a sparkline so you can see the shape of the period, not just its total.
| Metric | What it counts |
|---|---|
| Conversations | Conversations started in the period. A badge shows how many are live now |
| Active users | Distinct people seen in the period, identified or anonymous |
| Messages | User and agent messages combined |
| Run success rate | Completed runs as a share of runs that reached a terminal state |
Success rate deliberately ignores runs still in flight. Only completed, failed, and stopped runs count toward it, so an active period does not depress the number while messages are still streaming.
Conversation activity overlays message volume as a line on conversation counts as bars. The ratio between them is the interesting part: many conversations with few messages each suggests people are bouncing after one question, while the reverse suggests genuine back-and-forth.
Run performance plots completed against failed runs over time. A step change here usually lines up with a deploy, a prompt change, or an MCP server going unreachable.
| Metric | Meaning |
|---|---|
| Avg. time to first chunk | How long people wait before text starts appearing |
| Failed runs | Runs that ended with an error |
| Sandbox runs | Runs that reached a terminal state — completed, failed, or stopped |
| Credits used | Credits charged during the period |
Time to first chunk is the latency number that matters most to how the agent feels. Total latency can be long without anyone minding, as long as output starts quickly.
| Question | Page |
|---|---|
| Who is using the agent? | Users |
| What are people asking, and what did we say? | Conversations |
| Why did that fail, and what did it cost? | Logs |
Monitoring is scoped to one project. Use the project selector in the dashboard header to switch, and keep staging in a separate project so test traffic stays out of these numbers.
Every activity page shares the same controls: a search field, a date range with presets from Today through Last 30 days, and page-specific filters. Selections live in the URL, so a filtered view is a link you can share with a teammate.
Custom date ranges reach back three months. Results page 50 rows at a time with previous and next controls.
