Module 07 Summary

What this module established

Decide what to log by writing down the questions you will ask after a failure. Log resolved inputs rather than templates - the difference is where the surprises are. In a crew the visible error and its cause are usually in different agents, so read task outputs in order.

Carry forward

• Attribute tokens per agent. Routing overhead produces no visible output and is often the largest single line.
• Run two loop detectors: a hop ceiling for slow drift, a repeat count on (agent, action) pairs for tight loops. Each misses what the other catches.
• Emit one structured record per run, summarising outputs rather than storing them whole.

Before moving on

Move on when someone who was not present can answer all seven failure questions from your run record alone.

Mark complete