Agent failure modes — CCA-F Exam Prep

PencilPrepPencilPrep
L2.25|Agent failure modes
1/12
Real story
A server dashboard showing a single AI agent process that has been running for 14 hours. Token count: 2.3 million. API cost: $847. The agent is in iteration 2,941 of a loop, still trying to 'improve' a document that was good enough at iteration 3. A panicked engineer stares at the screen.

$847 in API costs. One agent. Fourteen hours.

A company deployed an AI agent to improve marketing copy. The agent's loop: write a draft, evaluate it, improve it, repeat. No maximum iteration count. No spending limit. No stop condition other than 'until it's perfect.'

The agent decided the copy was never perfect. Draft 3 was great. Draft 47 was identical to draft 46. Draft 2,941 was worse than draft 3. But the agent couldn't tell. It had no way to evaluate 'good enough.' So it kept going.

The agent wasn't broken. It was doing exactly what it was told. It just never had a reason to stop.