# Reflection

A feedback pattern in which a model critiques an attempt and records lessons or a revised approach for another attempt.

technical-term · Evaluation & verification · emerging · Reviewed 2026-08-09

## Definition

Reflection asks an agent or model to inspect an outcome, feedback, or trajectory and produce a critique, diagnosis, or revised plan. It can improve later attempts, but it remains correlated self-review unless a separate verifier, model lineage, deterministic tool, or accountable expert independently challenges the result.

Research and practical use show promise, while reliability, calibration, and independence remain task-dependent.

## Why it matters

Reflection creates a feedback artifact that can alter the next loop turn.

It should be treated as a builder improvement mechanism, not the final acceptance authority.

## System anatomy

- **Signal:** Test output, tool result, user correction, or other evidence.
- **Critique:** A diagnosis of why the attempt succeeded or failed.
- **Update:** A revised plan, memory item, or candidate output.



## Important distinctions

- **Independent verification:** Self-reflection shares model context and blind spots; independent verification separates judgment and block authority.
- **Retry:** Reflection should introduce new evidence or a changed strategy before another attempt.



## Implementation signals

- Ground critique in external evidence
- Keep builder reflection separate from release acceptance
- Evaluate whether reflections improve outcomes rather than merely sound persuasive

## Failure modes

- Confident rationalization of the original error
- Poisoned feedback becomes durable memory



## Related knowledge

- [Agent loop](https://aisdlc.ai/agentic-engineering/agent-loop) — The finite runtime cycle in which an agent interprets state, selects an action, observes the result, updates state, and decides what happens next.
- [Independent verifier systems](https://aisdlc.ai/agentic-engineering/multi-agent-verification) — A separation-of-judgment architecture in which builder agents, evaluator agents, deterministic checks, domain experts, and authorization authorities challenge different failure surfaces.
- [Eval-driven development](https://aisdlc.ai/agentic-engineering/eval-driven-development) — An engineering loop that converts expected behavior and observed failures into repeatable evaluations combining deterministic checks, environment inspection, security testing, model graders, repeated trials, and human judgment.
- [LLM as judge](https://aisdlc.ai/agentic-engineering/llm-as-judge) — The use of a language model to score, compare, classify, or critique another system’s output or trajectory.

## Sources and further study

- [arXiv / NeurIPS — Reflexion: Language Agents with Verbal Reinforcement Learning](https://arxiv.org/abs/2303.11366) — Introduces a feedback pattern in which an agent converts signals into verbal reflections retained for later attempts. **Use:** Primary definition. **Limitation:** Self-reflection can improve subsequent attempts in some tasks but is not independent verification and can preserve or amplify a model’s blind spots.
- [DeepLearning.AI — Agentic AI](https://www.deeplearning.ai/courses/agentic-ai/) — A practitioner course organizing agentic systems around reflection, tool use, planning, multi-agent workflows, evaluation, and optimization. **Use:** Official guidance. **Limitation:** Course taxonomy is instructional rather than normative; it does not itself define enterprise authorization or assurance requirements.
- [Anthropic — Demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) — A practical treatment of evaluating trajectories, outcomes, graders, tasks, and agent-environment interaction. **Use:** Official guidance. **Limitation:** This is first-party engineering guidance drawn from Anthropic deployments. Evaluation designs remain task- and environment-specific, and the article does not establish independent verifier ownership by itself.

---

This library synthesizes cited research, standards, official documentation, and clearly attributed practitioner perspectives. Maturity describes the state of a concept—not vendor endorsement, production readiness, or permission to deploy. Benchmarks and demonstrations do not replace use-case evaluation, governed controls, independent verification, or named human release authority.