ModelRefs / Critic Agent — AI Glossary

Critic Agent — AI Glossary

A secondary agent reviewing and providing feedback on another agent's output to improve quality through iterative refinement.

Overview

Critic agents implement a generator-critic loop: the generator agent produces a draft response or plan; the critic evaluates it against a rubric or performs fact-checking; the generator revises. Used in Reflexion, Constitutional AI self-critique, and code-review agents. Can use the same or a stronger model for the critic role.

Reference details

Topic	agents
Last reviewed	2026-06-24

How it works

A generator produces a first attempt against the task, with no expectation that it is final.
A critic receives the task and that attempt, and is prompted to evaluate rather than to improve — its output is a list of specific defects, not a rewrite.
The rubric matters more than the model: a critic asked for general feedback returns praise, while a critic asked to check named properties returns findings.
Those findings go back to the generator alongside the original attempt, so the revision is informed by what was wrong rather than regenerated from scratch.
The loop terminates on an empty critique, a fixed round limit, or a critic verdict, since an unbounded loop will keep finding things to change.
Separation is what makes it work: a model asked to critique its own output in one pass tends to defend it, while the same model in a distinct call with only the artefact in front of it does not.

Example: The loop costs three passes, which sets the decision rule

Price a single round of generate, critique, revise. The generation call takes 2,000 input and 800 output tokens. The critique call takes 2,800 input — the prompt plus the draft — and 400 output. The revision call takes 3,200 input, adding the critique, and 800 output. Totals are 2,000 + 2,800 + 3,200 = 8,000 input and 800 + 400 + 800 = 2,000 output tokens. At $3 per million input and $15 per million output that is 0.008 x $3 = $0.024 plus 0.002 x $15 = $0.030, so $0.054. The single-pass baseline is 0.002 x $3 = $0.006 plus 0.0008 x $15 = $0.012, so $0.018. The loop therefore costs $0.054 against $0.018, exactly 3 times a single pass. That ratio turns the architecture question into an arithmetic one. Suppose single-pass output carries a 12% defect rate and the loop cuts it to 4%. Over 100 tasks, single-pass costs 100 x $0.018 = $1.80 and leaves 12 defects; the loop costs 100 x $0.054 = $5.40 and leaves 4. So $3.60 of extra spend prevents 8 defects, which is 3.60 / 8 = $0.45 per defect avoided. The loop is worth running whenever handling a defect downstream — a support contact, a rollback, a human correction — costs more than $0.45, and is not when it costs less. That is a question about the defect's consequences, which is where the decision belongs.

Commonly confused with

A critic agent is not a reward model, though both score outputs. A reward model is trained on preference data and emits a scalar used for optimisation, usually during training and at enormous volume; a critic agent is a prompted model emitting natural-language findings at inference, used to drive one revision. The outputs are not interchangeable: a scalar cannot tell a generator what to fix. It is also distinct from self-critique within a single response, where the model reviews its own reasoning in the same generation. The separation into distinct calls is doing real work — a model continuing its own text is committed to it, while a model shown an artefact without having just produced it evaluates more honestly.

When to use it

Reach for it when:

Where a defect is expensive downstream, which the per-defect figure above makes a direct comparison.
On tasks with checkable properties, since a specific rubric is what separates findings from praise.
When the critique can be bounded to a round or two, before the cost multiple grows past the defect it prevents.
With generator and critic in separate calls, which is the arrangement that produces honest evaluation.

Reach for something else when:

On high-volume low-stakes output, where tripling cost to avert a cheap defect loses money.
Where no rubric exists, which yields agreeable commentary rather than findings.
In an unbounded loop, since a critic will always find something and the rounds are each a full pass.
As a substitute for a deterministic check, where a linter, a schema validator or a test suite is cheaper and certain.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Critic Agent — AI Glossary.

Frequently asked questions

What is Critic Agent?