Skip to main content
Atlan Trident builds a skill’s usage from traces: when an agent runs and its trace records the skill, Atlan Trident attributes the run to that skill and adds it to the Usage tab. Use it to see whether a skill is being used, by whom, and how its runs go.

Attribution

You do not record a separate usage event. When a runtime sends traces to Atlan Trident, Atlan Trident matches the skill a run used to the skill and, where it can, to the exact version. A run is attributed only when its trace carries enough information to identify the skill, so usage appears only for runtimes that report traces.

Usage tab

For the selected period, the tab shows:
  • Runs: Each time an agent used the skill end to end.
  • Active users and People reached: Distinct people who ran the skill.
  • Agents: Distinct coding agents that ran it.
  • Success rate and Error rate: Share of runs that finished without an error, and share that logged at least one.
  • Tokens per run and Tools: Average tokens a run consumes, and distinct tools the skill called.
Each figure shows the change from the previous period. A figure Atlan Trident cannot read for the period shows a dash, not a zero. Below, the run list can be filtered by agent and user, and each run links to its trace when you have permission to open it. When skill improvement is configured and enough recent evidence exists, editors can Improve skill from recorded runs. Skill quality then shows judgments about reaching the goal, choosing the right skill, and needing correction for the reviewed sample. These judgments are separate from the run list’s execution error rate.

Example

meeting-notes-to-actions shows 120 runs this month across 14 people and two coding agents, with a 96% success rate. One run with an error links to its trace, which shows the skill called a calendar tool that refused access, so the owner adds a prerequisite to SKILL.md.

What usage does not tell you

  • Whether results were correct. A high run count shows reuse, not quality. Use Evaluate agents to measure quality.
  • Untraced use. A runtime that does not report traces contributes nothing; No runs yet can mean missing instrumentation. See No traces for skill in use.
  • Every user’s runs in the list. The run list can be limited to your own runs while the figures count everyone’s.