--- schema: formation.pattern/v0.1 kind: pattern visibility: public aliases: - 'evaluator-optimizer (Anthropic)' - reflection - self-refine - 'generator-critic loop' - 'actor-critic prompting' canonical_url: https://topologyindex.com/patterns/critic_loop.md description: 'Implementation and criticism alternate for a bounded number of rounds.' executable_here: true executor_requirements: - sequential_role_invocations - declared_handoff_on_measured_signal - review_only_role - bounded_repeat_activation family: verification findings_page: /patterns/critic_loop/findings.md observed_in: [] path: /patterns/critic_loop.md pattern: critic_loop pattern_index: /patterns/index.md product_api_version: v1 published_findings: - benchmarks: - GSM8K compared_against: 'One-step generation with the same model' direction: helped source_id: arxiv:2303.17651 task_domain: 'Dialogue, code optimization, math reasoning and other generation tasks' task_domains: - coding - reasoning url: https://arxiv.org/abs/2303.17651 - benchmarks: - HumanEval - ALFWorld - HotpotQA compared_against: 'The same agent without reflection' direction: helped source_id: arxiv:2303.11366 task_domain: 'Sequential decision-making, coding and language reasoning' task_domains: - coding - reasoning url: https://arxiv.org/abs/2303.11366 - benchmarks: - TriviaQA - HotpotQA - GSM8K compared_against: 'The same model without tool-interactive critiquing' direction: helped source_id: arxiv:2305.11738 task_domain: 'Free-form question answering, math program synthesis and toxicity reduction' task_domains: - reasoning url: https://arxiv.org/abs/2305.11738 - benchmarks: - GSM8K - CommonSenseQA - HotpotQA compared_against: 'The model''s initial answers before intrinsic self-correction' direction: hurt source_id: arxiv:2310.01798 task_domain: 'Math, commonsense and multi-hop question answering' task_domains: - reasoning url: https://arxiv.org/abs/2310.01798 - benchmarks: - 'Game of 24' - 'Graph Coloring' - 'STRIPS planning' compared_against: 'Iterative prompting with a sound external verifier and one-shot generation' direction: hurt source_id: arxiv:2402.08115 task_domain: 'Reasoning and planning puzzles' task_domains: - reasoning url: https://arxiv.org/abs/2402.08115 - benchmarks: [] compared_against: 'Various self-correction baselines across surveyed studies' direction: mixed source_id: arxiv:2406.01297 task_domain: 'Survey of self-correction research' task_domains: [] url: https://arxiv.org/abs/2406.01297 - benchmarks: [] compared_against: 'The same idea generator with fewer or no rounds of reviewing-agent feedback' direction: helped source_id: arxiv:2404.07738 task_domain: 'Research idea generation over scientific literature' task_domains: - research url: https://arxiv.org/abs/2404.07738 qualifying_evidence: [] references: - https://arxiv.org/abs/2303.17651 - https://arxiv.org/abs/2303.11366 - https://www.anthropic.com/engineering/building-effective-agents schema_version: v0.1 title: 'When to use the critic_loop multi-agent pattern (evaluator-optimizer, reflection)?' unmet_executor_requirements: [] --- # When to use the critic_loop multi-agent pattern (evaluator-optimizer, reflection)? **Short answer:** start with it when the task needs several rounds of criticism; avoid it when rounds stop converging. A hypothesis from the [decision guide](/patterns/index.md), not a ranking; [published findings](/patterns/critic_loop/findings.md) keep the unfavourable ones. The `critic_loop` multi-agent pattern (family `verification`): implementation and criticism alternate for a bounded number of rounds. Also called: evaluator-optimizer (Anthropic), reflection, self-refine, generator-critic loop, actor-critic prompting. ## The arrangement - Structure: The implementing and review roles, with a declared handoff in each direction. - Control: Both handoffs are declared and each carries its own activation limit. - Shared state: The episode workspace, carried across rounds. ## When it is hypothesised to fit Hypothesised to fit when one round of criticism is rarely enough and each further round is expected to remove more than it costs. A hypothesis to test on a workload, not a finding: this service has not tested it. ## Failure modes to watch for - Rounds that repeat without converging, which is why the activation bound exists and why crossing it is recorded. - The loop can consume the budget before the validator ever runs. - Later rounds can reopen questions the earlier rounds settled. These are things to watch for, not outcomes anyone measured here. ## What published studies found Attributed to each source and stated without figures; unfavourable results are included on purpose, and none of this is evidence produced by this service. What each was compared with is in `published_findings`; each finding in words, with its caveat, is at [/patterns/critic_loop/findings.md](/patterns/critic_loop/findings.md). Reviewed 2026-09-23. Each label says how this pattern fared against what it was compared with, as the source reports it; the labels are defined at [/docs/schemas/pattern/v0.1.md](/docs/schemas/pattern/v0.1.md). - **Helped**: [Self-Refine: Iterative Refinement with Self-Feedback](https://arxiv.org/abs/2303.17651) - **Helped**: [Reflexion: Language Agents with Verbal Reinforcement Learning](https://arxiv.org/abs/2303.11366) - **Helped**: [CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing](https://arxiv.org/abs/2305.11738) - **Hurt**: [Large Language Models Cannot Self-Correct Reasoning Yet](https://arxiv.org/abs/2310.01798) - **Hurt**: [On the Self-Verification Limitations of Large Language Models on Reasoning and Planning Tasks](https://arxiv.org/abs/2402.08115) - **Mixed**: [When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs](https://arxiv.org/abs/2406.01297) - **Helped**: [ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models](https://arxiv.org/abs/2404.07738) ## Sources that describe it Where the arrangement is described. A source is a reason to test it, never a result for it. - [Self-Refine: Iterative Refinement with Self-Feedback (Madaan et al.)](https://arxiv.org/abs/2303.17651) - [Reflexion: Language Agents with Verbal Reinforcement Learning (Shinn et al.)](https://arxiv.org/abs/2303.11366) - [Building effective agents (Anthropic)](https://www.anthropic.com/engineering/building-effective-agents) ## Can this deployment execute it Yes. The runner provides everything this arrangement needs: - `sequential_role_invocations` (one role invocation at a time, in a declared order): The supervisor invokes exactly one role at a time and ends on the first crossed bound. - `declared_handoff_on_measured_signal` (a handoff whose condition is a measured fact about the last invocation): Handoff conditions are evaluated from trusted instrumentation, never from role output. - `review_only_role` (a role that can read the workspace but not write it): A role without the write tool is a review-only role, and its completion is a signal. - `bounded_repeat_activation` (a handoff that may fire more than once under a declared bound): Each handoff carries its own activation limit; crossing it is a recorded observation. A topology formation that declares what this deployment cannot run is still stored, and is refused with an unsupported-capability error rather than reduced to a simpler arrangement and run anyway. ## Evidence for this pattern None. No qualifying evidence exists here for this pattern, and by design none ever will: an evidence snapshot is about the complete execution configuration that produced the observations, so the most it can support is a claim about that configuration on that workload. ## Observed in No research record on this service reports observing this arrangement. ## How evidence works here A pattern names how work is organized, never how well that works. Evidence in this service attaches to a complete execution configuration — a topology formation together with an execution lock pinning models, tools and runtime — and never to a pattern on its own, so no page here carries a success rate, a cost or a run count for an arrangement.