Evaluation & Operations
Common AI Agent Failure Modes
A field guide to loops, premature success, bad tool calls, stale state, weak evidence, and silent escalation failures.
Action failures
Agents may choose the wrong tool, invent an argument, repeat a call, or execute steps in an unsafe order. Narrow schemas and state-aware validation reduce the blast radius.
Reasoning failures
An agent can anchor on an early assumption, ignore contradictory observations, or mistake fluent text for evidence. Verification steps should inspect external state rather than ask the same model if it is correct.
Control failures
Missing limits lead to endless loops, excessive spending, or unauthorized actions. Time, step, cost, and permission boundaries must be visible to both the runtime and operators.