Agent failure modes — CCA-F Exam Prep

L2.25✨ 0

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.

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. It was still trying to 'improve' a document that was perfect at iteration 3.

A company deployed an agent to improve marketing copy. Loop: write a draft, evaluate it, improve it, repeat. No maximum iterations. No spending limit. No stop condition except 'until it's perfect.' Draft 3 was great. Draft 2,941 was worse. The agent couldn't tell the difference.