Aglet

Turn Instruction Failures Into Better Agent Evaluation

Instruction cases show where evaluation confuses fluency with compliance. Preserve the request, obligation map, actual behavior, and user effect. Turn the lesson into a reusable case that makes priorities, exclusions, and conditional paths visible in future reviews of the same type.

Keep the lesson for the next incident

  1. Save the instruction case

    Store request, context, obligation map, output, tool actions, satisfied and missed requirements, branch, format, score, and interpretation. Include a case where concise output passed and extra text would have violated an exclusion.

  2. Improve compliance practice

    Add explicit priority, condition, exclusion, format, and ambiguity checks to evaluation plans. Require a matched conflicting request. Explain how the revision addresses the exact silent miss or misprioritization observed in review.

  3. Watch for fluent violations

    Set a signal such as outputs passing broad quality review while missing hard constraints, conditional branches never tested, formats failing downstream, or exclusions ignored in long answers. Assign an owner to sample obligation maps and revise checks when it appears.

What to carry forward

The learning record should connect the missed obligation to the revised contract or case practice and recurrence signal. State which priority or branch became explicit. Keep the lesson tied to the tested request family rather than treating verbosity or tone as compliance.

Technical background: OpenAI evaluation guidance.

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