Aglet

Verify an Agent Trajectory Against State

Verification makes the path reviewable without rewarding one golden sequence. Apply explicit transition checks to direct and recovery cases, then inspect exceptions. A passing trajectory should reach the intended state through an allowed path with evidence for each material transition.

Check whether the outcome improved

  1. Set transition checks

    Require precondition, observation, call, argument, return, state, dependency, retry, and completion evidence where applicable for each run. Mark acceptable alternate paths. Define how missing traces and external failures are recorded.

  2. Review path samples

    Apply checks to direct success, chained work, failed tools, retries, loops, partial completion, and valid alternatives. Compare snapshots with calls and returns. Record whether a correct final answer followed an invalid or unobserved transition.

  3. Approve the trace boundary

    State which path behaviors the result supports and what needs deeper planning or outcome review. Record fixture, versions, reviewer, and retest trigger. Keep a result inconclusive when a material transition cannot be observed.

What to carry forward

The verification record should show transition rules, path samples, state evidence, alternate paths, and limitations. Approve only the trajectory behavior observed. Reopen it when tools, state representation, recovery rules, or task boundaries change before using the score to compare agents when another reviewer repeats the path.

Technical background: Agent evaluation research on arXiv.

Keep the decision with the work.

Use a Work Item in Aglet to record the problem, the evidence you have, and the next decision. Add an owner and priority, then keep updates in the discussion so the next person can follow the reasoning.

Create an account See the product workflow