Evaluate an organization
Freeze sanitized evidence against the public AgentOS benchmark before improving it.
Use the evaluation Skill to measure Quickstart, Fleet delivery, composition integrity, recovery, human attention, efficiency, robustness, safety or accountability.
Evaluation selects a scenario, pins the exact subject and environment, observes the declared run and produces a sanitized evidence bundle. It records failed and incomplete attempts, unknown telemetry and limitations. It does not change the organization while measuring it.
Why freeze first
Suppose a Fleet stalls after its useful supervision waits are consumed. Repairing the Skill during the run would erase the evidence needed to distinguish the original failure from the improved behavior. Instead:
- end and classify the run;
- sanitize and freeze the complete evidence bundle;
- verify its schema and chain of custody;
- begin a separate improvement review.
The benchmark is public and portable. It can evaluate an Agent organization that does not use AgentOS, provided the evidence satisfies the same contract.
Last updated on