Module 07 Activity

Scenario

Something in the crew produced a wrong answer without raising. This activity is about finding it in minutes.

What you build

A per-task trace, per-agent cost attribution, two loop detectors, and one structured run record.

Steps

1. Write down the questions you will ask after a failure, and the field each one needs. Log resolved inputs, not templates.
2. Plant an error in an early task and find it by reading task outputs in order. Note whether the visible symptom was in the same agent.
3. Attribute tokens per agent per task and report the shares. Identify the largest line that produces no user-visible output.
4. Implement both loop detectors - a hop ceiling and a repeat count on (agent, action) pairs - and show a case each one catches alone.
5. Emit one structured run record per run, with output summaries rather than full outputs.
6. Give the record to someone who was not there and check they can answer every question from step 1.

Evidence to hand in

• The question-to-field table.
• The planted error, where the symptom appeared, and where the cause was.
• The per-agent cost shares.
• Both loop detectors with a case each catches alone.
• One run record, and the outcome of the reader test.

Review checklist

• Resolved inputs are logged, not templates.
• The planted error demonstrates symptom and cause in different agents, or explains why not.
• Cost is reported as shares, not only dollars.
• Both detectors exist and each has a case the other misses.
• A reader who was not there answered every question.

Mark complete