| name | agent-behavior-audit |
| description | Audit whether an autonomous agent actually observes state, uses tools, improves from outcomes, and stays aligned to user intent using traces, logs, and artifacts. |
Agent Behavior Audit
Use this when product claims about autonomy need proof.
Treat UI labels, docs, and architecture diagrams as claims until traces or tool records prove them.
Flow
- Identify the promised agent behavior and the real user task.
- Collect traces, logs, tool calls, state changes, prompts, outputs, and run artifacts.
- Check whether the agent observed real state before acting.
- Check whether it used real tools or only narrated actions.
- Check whether it reflected on outcomes and changed behavior across turns/runs.
- Compare actual behavior to user intent and product claims.
- Classify gaps as fake action, blind action, no learning, bad incentives, or missing observability.
Output
Report claim, evidence, verdict, impact, and required fix for each behavior gap.
Use references/full-reference.md for the full question list and red flags.
Then consider
harden if fake or unsafe behavior creates a security risk.eval-agent if the product needs a repeatable behavior evaluator.
