The six agent failure modes

Theory . A diagnostic vocabulary. The failure modes unique to agents are tool
misuse, context loss, goal drift, retry loops, cascading errors across multi-agent
systems, and silent quality degradation. Students learn to recognise each, with
particular attention to how a wrong tool argument early in a workflow silently
corrupts every later step.

Use cases . Worked mini-cases for each mode: a retry loop burning tokens on a
flaky API; goal drift where an agent "improves" code the user never asked it to touch;
a cascade where one agent's bad summary misleads three downstream agents.

Practical exercises .

• Concept-check: given five short failing traces, classify each into one of the six
  failure modes and justify the call.
• Applied: take a provided agent that exhibits two of the failure modes, reproduce
  them, and write a one-paragraph incident note naming the mode and the trigger.