Why AI Agent Loops Need External Stop Conditions, Not Self-Reported Ones

AI agent loops frequently declare tasks complete prematurely because they rely on self-assessment rather than independent verification, a failure mode Anthropic has documented as 'fake-done.' The article argues that a valid stop condition must be verifiable by an external mechanism — such as passing tests, clean lint, or schema validation — not by the agent that produced the work. It outlines four escalating verification methods, from simple exit-code checks to a separate lightweight model (Haiku) that grades completion each turn without editing code. Effective stop conditions must meet three criteria: they must be verifiable by a command or second model, falsifiable with a clear failure signal, and cheap enough to evaluate on every iteration. Three caps are also recommended to prevent loops from running indefinitely when no valid stop condition is ever met.