Aglet

Turn Trajectory Results Into Better Agent Checks

Trajectory reviews show where outcome-only scoring loses the reason a task succeeded or failed. Preserve the first divergence, state evidence, recovery path, and decision impact. Turn that case into a reusable trace check that supports diagnosis without prescribing one sequence.

Keep the lesson for the next incident

  1. Save the trajectory case

    Store task fixture, ordered states, observations, calls, returns, retries, side effects, final answer, score, and interpretation. Include a successful outcome with a faulty intermediate transition and a failure caused by missing evidence.

  2. Improve trace practice

    Add state snapshots, transition labels, alternate-path rules, recovery checks, and missing-evidence statuses to evaluation plans. Explain how the change addresses the first divergence rather than merely adding more trace volume.

  3. Watch for path blind spots

    Set a signal such as final answers passing while state checks fail, repeated retries without snapshots, or loops counted as success. Assign an owner to sample trajectories and revisit the check when the signal appears.

What to carry forward

The learning record should link the first divergent state to the revised trace practice and recurrence signal. State which transition evidence became required. Keep the lesson tied to the task contract and observed states instead of treating path length as quality over time.

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