What is a generator-critic loop?
A generator-critic loop, also called evaluator-optimizer, is an agent pattern in which one model produces a draft, a separate model grades it against set criteria, and the draft is revised until it passes or a limit is reached.
Anthropic's "Building effective agents" (December 2024) called this pattern evaluator-optimizer and gave two signs that it fits: responses get better when a person explains what is wrong, and a model can give that same feedback itself. Literary translation and multi-round search were its examples.
The case for a separate critic got stronger in 2026. Anthropic's post on harness design for long-running apps (March 2026) observed that agents asked to judge their own output tend to praise it, even when it is mediocre, and that a standalone evaluator is easier to tune toward scepticism than a generator is to make self-critical.
Three names, one pattern
Anthropic's 2024 guide says evaluator-optimizer. Google's guide to multi-agent patterns in ADK (December 2025) says generator and critic for a pass or fail quality gate, and iterative refinement for a loop that polishes a draft, both built on a LoopAgent with a max_iterations cap. Anthropic's 2026 harness post describes a generator and evaluator inspired by generative adversarial networks. The structure is the same in each.
What a working critic needs
Explicit criteria
A rubric the critic scores against. In Anthropic's frontend example these were design quality, originality, craft and functionality, with hard thresholds that fail the round.
Tools to check, not just read
The same example gave the evaluator a browser through Playwright so it could click through the running app and take screenshots.
Calibration
Few-shot examples with detailed score breakdowns kept the evaluator's judgement steady across runs.
An agreed definition of done
A sprint contract, negotiated between generator and evaluator before work starts, fixes what will be checked.
A cap
A maximum number of rounds, so a draft that cannot pass stops costing money.
What it costs
In Anthropic's harness post, a retro game maker built by a single agent took 20 minutes and cost $9; the full harness with planner, generator and evaluator ran for 6 hours and cost $200, with much higher quality. Generations went through 5 to 15 critique rounds. With a later model, the evaluator was cut to a single pass at the end and kept only where the task was beyond what the model handled reliably alone.
How Motiontic's engineers apply it
Motiontic's engineers add a critic where the output has checkable criteria and errors are costly, such as generated documents, code changes or customer replies. The critic's rubric is versioned and checked against human grades, and the loop is removed for any task where evals show the generator passes on its own.
Self-critique compared with a separate critic
| Self-critique | Separate critic | |
|---|---|---|
| Who grades the draft | The model that wrote it | A different prompt, model or agent |
| Bias | Tends to rate its own work too highly | Can be tuned toward scepticism |
| Tools | Usually none beyond the draft | Can run tests, browsers or other checks |
| Cost | Lower | Higher, one extra call or more per round |
| Fits | Quick fixes to formatting and obvious slips | Work with clear criteria and costly errors |
Common questions
What is the difference between a generator-critic loop and LLM as a judge?
LLM as a judge is the grading component: a model scoring output against a rubric. A generator-critic loop feeds that grade back to the generator so it can revise, and repeats.
Is a generator-critic loop the same as a GAN?
It borrows the idea of a generator and a discriminator from generative adversarial networks, but nothing is trained. The critic's feedback changes the next draft, not the model's weights.
How many rounds should the loop run?
Set a cap from evals. Anthropic's harness runs used 5 to 15 rounds per generation, and Google's ADK pattern enforces a hard limit with max_iterations.
When is a generator-critic loop not worth it?
When there are no clear criteria to grade against, or when evals show the generator already passes without feedback. In both cases the critic adds cost without adding quality.
Related
LLM as a judge
A model grading model output against a rubric, checked against human labels.
Eval
Repeatable, scored tests that tell a team whether an AI change made things better or worse.
Long-running agent
Agents that work for hours or days, and the state that lets them pick up where they stopped.
Agentic workflow
Model calls on a path set in code: predictable, cheaper and easier to test than a free-running agent.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Self-reflection (agents)
An agent critiques its own attempt in words and tries again with that critique in hand.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.