# Reflection Loop — Self-Review Mechanism

> An agent reviewing and critiquing its own output before finalizing it — generate, examine, revise, repeat. The reflection loop builds a draft-and-review discipline into the system itself, catching errors at the moment they're cheapest to fix: before anyone else sees them.

**Canonical URL:** https://www.andekian.com/ai-lexicon/reflection-loop  
**Author / Site:** Stephen Andekian — https://www.andekian.com

**Term 79 of 100** · Agentic Systems  
**Tags:** Self-Critique, Revision, Quality, Draft-Review

## Key Stats

- **Pattern — draft → critique → revise:** The writing-process discipline, systematized — first attempts treated as drafts by architecture.
- **Lift — consistent:** Reflection measurably improves accuracy, completeness, and constraint adherence across task families — one of agentic AI's most reliable patterns.
- **Cost — 2–3x tokens:** Critique and revision passes multiply per-output compute — quality bought with latency and spend, tiered to stakes.

## What Reflection Loop Actually Is

First drafts are first drafts, whether human or machine — and single-pass generation amounts to shipping them. The reflection loop adds what good work always had: a review step. The system generates, then examines its own output — against the task's requirements, against the evidence, against explicit quality criteria — produces a critique, and revises. The error caught in reflection costs tokens; the same error caught downstream costs trust, rework, or consequences.

The mechanism works because critique is easier than creation. Models detect flaws in presented work more reliably than they avoid those flaws while generating — the same asymmetry that makes human editors valuable. Effective reflection sharpens this further with explicit criteria: not “improve this” but “check every figure against the source,” “verify each requirement is addressed,” “list claims lacking citations.” Vague reflection produces cosmetic revision; targeted reflection produces corrections.

Variants tune the pattern to stakes. Single-pass reflection catches the worst cheaply; iterative loops repeat until criteria pass or budgets cap; separate critic configurations — a fresh model instance unanchored to the draft's reasoning, sometimes a different model entirely — strengthen the review by removing the author's bias toward their own work. In agent workflows, reflection extends beyond outputs to actions: plans reviewed before execution, results examined after, the loop standing in wherever a careful human would pause and check.

The pattern's limits deserve clear eyes. Reflection inherits the reviewer's blind spots — a model that doesn't know a fact is wrong won't flag it; self-review without external grounding polishes more than it corrects. Convergence isn't guaranteed: loops can oscillate or drift past their best draft, which is why iteration caps and pass-fail criteria matter. And the economics are real: two to three times the tokens per output buys the quality lift — a price worth paying at consequence and worth skipping for the disposable. Tier the loop to the stakes, like every other reliability layer.

## How It Works: Building review into the loop

Reflection inserts a critic between generation and delivery — the output examined against criteria, revised against findings, and only then released.

1. **Generation** — The first attempt produces — treated by the architecture as a draft, not a deliverable.
2. **Criteria Recall** — The review's standards assemble — task requirements, evidence checks, quality rubrics — the lens for examination.
3. **Critique** — The draft is examined against criteria — flaws, gaps, and unsupported claims surfaced as explicit findings.
4. **Revision** — The draft updates against the findings — corrections made, gaps filled, claims grounded or removed.
5. **Convergence Check** — Criteria pass, or the loop repeats within its budget — iteration bounded by caps and pass-fail gates.
6. **Release** — The reviewed output delivers — with critique history logged as the quality trail of how it got there.

## Anatomy: The Components Teams Must Understand

- **Critic Role** (The internal reviewer): The configuration examining the draft — same model re-prompted, fresh instance, or separate model, in rising independence.
- **Review Criteria** (The explicit lens): Concrete checks replacing vague improvement — the specificity that turns reflection from cosmetic to corrective.
- **Critique Artifact** (Findings, recorded): The explicit list of flaws and gaps — actionable input for revision and a logged quality trail afterward.
- **Revision Pass** (Findings applied): The draft corrected against the critique — the step where review becomes improvement.
- **Iteration Bounds** (Convergence discipline): Caps and pass-fail gates preventing oscillation and budget burn — loops that end, by design.
- **External Anchors** (Grounding the review): Sources, tests, and validators feeding the critique — the outside truth self-review alone can't supply.

## Strategic Implications

- **Cheap insurance at the moment of creation** (01 · Quality): Reflection catches errors where they're cheapest — before delivery — and lifts accuracy, completeness, and constraint adherence across task families. For any output with consequences, the 2–3x token cost is among the best-priced reliability available.
- **Criteria make or break the loop** (02 · Design): Vague self-review polishes; explicit checks correct. Invest in the rubric — verifiable criteria, evidence requirements, pass-fail gates — and strengthen independence where stakes rise: fresh critic instances see what authors don't.
- **Reflection isn't verification** (03 · Limits): Self-review inherits the model's blind spots — unknown errors stay unflagged, and polish can masquerade as correction. Anchor critiques in external truth (sources, tests, validators) and keep reflection one layer in the reliability stack, not the stack itself.

## Common Misconceptions

- **Myth:** “Models can't meaningfully review their own work.”  
  **Reality:** Critique is empirically easier than creation — models catch flaws in presented drafts they failed to avoid while writing. The asymmetry is the pattern's foundation, and the measured lifts are consistent.
- **Myth:** “More reflection iterations mean better output.”  
  **Reality:** Returns concentrate in the first pass or two — beyond, loops oscillate, drift past their best draft, and burn budget. Bounded iteration with pass-fail criteria captures the value and skips the pathology.
- **Myth:** “Reflection makes external review unnecessary.”  
  **Reality:** Self-review shares the author's blind spots — it complements grounding, verification, and human gates rather than replacing them. The loop is one reviewer on the committee, not the committee.

## Related Terms

- [Chain of Thought — Sequential Reasoning Engine](https://www.andekian.com/ai-lexicon/chain-of-thought)
- [Agentic AI — Autonomous Workflow Execution](https://www.andekian.com/ai-lexicon/agentic-ai)
- [AI Agent — Autonomous AI Operator](https://www.andekian.com/ai-lexicon/ai-agent)
- [Recursive Reasoning — Multi-Pass Problem Solving](https://www.andekian.com/ai-lexicon/recursive-reasoning)
- [Self-Correction — Autonomous Error Fixing](https://www.andekian.com/ai-lexicon/self-correction)
- [Chain-of-Verification — Step-By-Step Validation](https://www.andekian.com/ai-lexicon/chain-of-verification)
- [Autonomous Execution — Reduced Human Intervention](https://www.andekian.com/ai-lexicon/autonomous-execution)
- [Observability — Production AI Monitoring](https://www.andekian.com/ai-lexicon/observability)

## Explore the Full Lexicon

All 100 terms: https://www.andekian.com/ai-lexicon

## Contact

Book a conversation or send an inquiry: https://www.andekian.com/#contact
LinkedIn: https://www.linkedin.com/in/andekian/