Verifier Loop

Simple Definition

A verifier loop is a feedback loop where an AI agent or an external system checks whether generated work behaves as expected. The check usually comes from tests, evals, CI, or other validation steps, not just the model’s own opinion.

How It Works

1. The agent produces some work (for example, a code fix).
2. A verifier runs, tests, a build, an eval, or a validation step.
3. The agent reads the result, sees what passed or failed, and adjusts.
4. The loop repeats until the work is clean, or the agent asks a human for help.

Example

A coding agent writes a fix, runs the test suite, reads the failure, updates the fix, and repeats, then asks for human review once the tests pass or when it gets stuck.

An Important Limit

A verifier loop reduces obvious errors, but it should not let AI approve its own work for high-risk changes. Tests can pass while the behavior is still wrong. Keep a human in the loop for anything that touches security, payments, or production, and review by risk.

Related Terms

• AI Loop, the general idea of an agent iterating in a loop
• AI Code Risk Review, deciding which results need human eyes
• Human-in-the-Loop, keeping people in control of approval
• AI Evals, structured ways to measure AI output quality