ai-automation-consultant / Failure Lab
Agent Runaway Loop Cost Cap
A coding agent repeatedly rewrites tests until the token spend explodes.
Category
agent-operations
Difficulty
advanced
Last Run
none
Incident Brief
Diagnose the agent failure and name the operating control that would have contained it.
Broken State
The agent was launched without a spend cap or bounded acceptance criteria.
Symptoms
- The task runs for hours and keeps changing unrelated assertions.
- Cost logs exceed the intended review budget.
Trace
agent.iteration=38 action=rewrite_tests cost_usd=18.42 task.acceptance_criteria=null spend_cap=null