Skip to main content
The Agent usage report compares every agent that recorded runs in a period, with a fleet summary, a trend chart, and a sortable agent table. Use it to see where spend and errors come from and to pick the agent to investigate. To investigate one agent you already know, open its profile’s Usage tab instead; see How trace inspection works.

Before you begin

Before you start, make sure you have:
  • Permission to manage organization settings: Usually the Admin role. Reports does not appear in the app rail for other members.
  • Traced agents: Agents whose runs are sent with their own API keys. See Send and verify traces.
1

Open Agent usage

Select Reports in the app rail, then Agent usage, and choose the workspace scope and period.Agent usage with a fleet summary, spend trend, and agent tableReports also holds AI usage, Shadow AI, Skill usage, Skill security, Repositories, and the AI ROI report. Shadow AI shows the MCP activity observed in coding sessions running the Atlan plugin, including servers reached outside MCP Gateway. It does not see tool calls from uninstrumented clients. Skill usage shows which registered skills people and agents call and how those runs go. Hovering Skills used opens a card with the share of registered skills used at least once in the period — a bar with Used in this period and Not used counts. Repositories shows agent usage by Git repository. A session counts for a repository when the Atlan CLI reports the Git remote it ran in, which Claude Code and Codex sessions run inside a checkout do, and a remote the registry does not hold yet can be registered from its row. Hovering Sessions or Recorded cost opens a card with the day-by-day trend for the period and the figure compared with the period before it. Selecting either one plots that measure on the Busiest repositories by chart; the picker beside the chart’s title switches the same measure. The AI ROI report shows what coding agents spend, what work they do, and how that work ends — its judged answers cover a sample of sessions, scored when you generate the report; see Run the AI ROI report.
2

Read fleet summary

What changed leads with the period’s token total, compared with the period before it, then shows spend, runs, active agents, span error rate, and cost per run for the period, each with its change from the previous one. Hovering Agent runs or Spend shows the measure day by day for the period, with what it covers; the card opens on keyboard focus too. Hovering Runs using skills opens a different card: it states how many of the period’s runs used a skill, with a Used a skill and No skill bar and each group’s count and share of all runs. Read it before comparing agents. When the tiles show their trend lines, a small legend beside the heading names the window they cover, such as Last 30 days.Switch the trend chart between Spend, Tokens, Runs, Cost / run, Tokens / run, and Span error rate. On a day-by-day chart with at least two days of data, Spend and Tokens stack each day’s total by the period’s top models, ranked by the metric shown, with the remainder as Other. Up to seven models are named, and the legend below the chart lists them, with Other drawn only when the remainder is not zero.The trend ends at the last whole day: a day still in progress is drawn only when the chart would otherwise hold fewer than three points, so on a two-day chart today appears. Totals elsewhere in the report, including What changed, still include today. The chart names day and week buckets in UTC: west of UTC, the day still in progress can read a day ahead of your local date.
3

Find agent that needs attention

Sort the agent table by spend, run count, cost per run, tokens per run, span error rate, or P90 latency, or search by name. Select a row to open the agent’s profile and inspect its Usage and recent runs.
4

See who agents ran for

The Visitors section shows who the fleet’s runs were for: the distinct people and calling agents named on runs by Identify visitors. Read the four figures — Visitors, Runs for visitors, Returning (visitors also present in the previous period), and Agents with visitors — then the cards below them: By agent, which agents ran for the most visitors, and Top visitors, which visitors got the most runs. Names come from the visitor directory; a visitor it has no row for shows by its handle.Filter the section with the Show control: Humans and agents, Humans, or Calling agents.If no run in the period named a visitor, the section shows a setup card instead of figures, with a prompt to copy into a coding agent opened in the busiest agent’s repository. The prompt walks through Identify visitors. A figure can also show as unavailable when the visitor directory or the period’s breakdown could not be fully read; a very long period can be incomplete, and a shorter one fixes it.See How trace attribution works for how a visitor attaches to a run.
5

Read the breakdown cards

The report also has six cards that break the period down: Model mix, spend, tokens, and cache efficiency for the six highest-spend models; Prompt cache, how much model input was reused instead of processed again; Tool performance, calls, fleet share, and recorded step errors for the six most-used tools; Top skills, invocation volume and share for the six most-used published skills; AI tool performance, run volume, spend, efficiency, and recorded errors by serving runtime; and Latency outliers, P50, P90, and tail gap for the six slowest agents. Human waits can be intentional.Each card’s title row shows a count of what the card covers and the number of days in the period, such as 6 models · 7 days on Model mix. A count that could not be read shows as a dash, not a zero.
6

Verify figures before acting

  • Span error rate counts spans recorded as errors. It does not measure response quality.
  • Cost / run and Tokens / run are period averages. Compare them with run volume before treating a change as a trend.
  • of all runs used skills is the share of runs with recorded skill use. Hovering the tile shows both counts — runs that used a skill and runs that did not — with their shares. A missing value means the breakdown could not be established.
  • A value marked unavailable could not be read; the report never turns it into a zero.
Each run’s cost comes from the token counts on its trace and your organization’s rates. See How cost is calculated.

Troubleshooting

If an agent you run is missing from the report, its runs were probably sent with a personal API key. See Traces appear under person instead of agent.

Next steps

See also