Reflection

What is reflection in an AI agent?

Reflection is the process through which an AI agent evaluates its progress, actions, or output and decides whether something should be corrected. It creates a feedback step between doing work and continuing.

A practical reflection loop is:

Act → Observe result → Evaluate → Identify problems → Revise strategy or output → Retry or continue

Where reflection happens

An agent can reflect after a tool result, failed action, completed section, or entire task. It may ask whether the result satisfies the goal, whether evidence is missing, whether an assumption failed, or whether another approach is needed.

For example, after drafting a report, an agent may compare it with the requested sections, notice that one claim lacks support, retrieve a source, and revise the draft.

Reflection versus reasoning

Reasoning is the broader process of interpreting information and making decisions. Reflection is reasoning focused on reviewing prior work or current progress.

Reflection also differs from planning. Planning organizes future work. Reflection evaluates what has already happened and may trigger replanning.

Evaluation is the measurement or judgment step. Reflection uses evaluation results to decide how to improve. A system can evaluate an answer without automatically revising it.

Reflection versus retrying

Retrying repeats an operation. Reflection first identifies why the previous attempt was insufficient and whether a different strategy is needed. Repeating the same failed action without new information is not meaningful reflection.

Reflection is also not memory, although stored episodes and feedback can provide evidence for later reflection.

Limits and failure modes

A model can confidently approve its own incorrect work or invent problems that are not present. Repeated reflection can increase latency, cost, and output drift. It may also create endless loops when success criteria are vague.

Reliable systems use explicit rubrics, tool-based verification, retry limits, and stopping conditions. High-impact outputs may still require independent review or human approval.

Why it matters

Reflection helps agents catch mistakes, respond to feedback, and improve multi-step work. It is most useful when outputs can be checked against clear criteria or external results.

The goal is selective review at meaningful checkpoints, not constant self-critique. See [Reflection in AI Agents](/reflection-in-ai-agents/) for the full relationship among evaluation, retrying, memory, planning, and correction.