Critic Agent

Simple Definition

A Critic Agent is an independent AI agent whose job is to evaluate another agent’s output, not to create the original work itself.

What a Critic Checks

• Correctness, whether the output is factually or logically right
• Completeness, whether it actually covers what was asked for
• Adherence to requirements, whether it follows the stated constraints
• Security, whether it introduces risk the builder may have missed
• Quality, whether it meets the bar, not just whether it technically runs
• Evidence, whether claims in the output are actually backed by something
• Usability, whether the result is actually usable as delivered

Why Separate Context Helps

A model reviewing the work it just produced tends to inherit the same blind spots and assumptions that shaped the original output. A critic agent running in a fresh context, or built on a different model, has no attachment to the first draft and no shared context nudging it toward agreement, which tends to produce a more honest, more independent review than asking the same pass to grade its own homework.

An Important Limit

A critic agent’s approval is not proof that the output is correct. It’s a stronger check than no review at all, but the critic can share the same weaknesses as any AI system: confident wrong answers, missed edge cases, and blind spots of its own. For anything with real consequences, a critic agent’s sign-off should support human judgment, not replace it.

Related Terms

• Gauntlet Loop, the builder-critic workflow a critic agent typically runs inside
• Verification Gate, the checkpoint a critic agent’s judgment feeds into
• Red/Blue/Green Review, a related three-stage review with more specialized roles
• Human-in-the-Loop, the check that should still sit above a critic agent for high-stakes work
• Multi-Agent System, the broader pattern of separate agents playing separate roles