Observability
Observability tracks historical usage, cost, and reliability across runs visible to you. Open it from Observability in the app.
This page aggregates recorded historical data. It is not a live activity feed and it does not replace the run session’s Trace tab.
Select a range
Choose one of these ranges:
- 24h
- 7d
- 30d
- 90d
The 24h range uses hourly points. Longer ranges use daily points. Changing the range reloads the historical summary.
Read the totals
The summary cards show totals for the selected range:
| Card | What it counts |
|---|---|
| Runs | Recorded runs. |
| Run cost | Generation-attributed cost in USD. |
| Run tokens | Generation-attributed input and output tokens. |
| Tool calls | Recorded tool calls. |
| Subagents | Recorded subagent tasks. |
| Errors | Tool errors, subagent failures, and model failures. |
Use the totals to identify costly or unreliable periods, then use Agent Ops or an individual run to investigate the relevant work.
Charts and breakdowns
The page charts the selected range for:
- Cost
- Tokens
- Tools
- Subagents
- Compactions
- Errors
Hover a chart point to see its value. Each chart also includes View accessible data, which exposes the same values as a table.
The bottom breakdowns list up to five entries each:
- Top tools, by call count.
- Subagent roles, by task count.
- Models, by cost, with input and output token detail.
Partial historical coverage
The page can show Partial historical coverage when some sessions do not have complete metrics or activity. Open that section to read the coverage warnings.
Treat totals and charts as the recorded subset when this warning is present. Missing coverage is not evidence that no work, cost, or error occurred.
Empty and unavailable data
No historical metrics were recorded in this range means there is no recorded data for the selected range. If the page cannot load metrics, it shows an error and Retry. If a refresh later fails, the page can continue to show the last successfully loaded data.
For a single run’s spans, timing, and task usage, open that run and use Trace. For operator-level Meta-Harness archives, see Meta-Harness trace capture.