Reflection Loops

Introduction

Reflection is the agent critiquing its own output and trying again if it's not good enough. Patterns: Reflexion, Self-Refine, Critic agents. Reflection often boosts quality 10-30% at the cost of more LLM calls.

Beginner analogy: writers don't ship first drafts — they read, criticise, rewrite. Reflection makes agents do the same.

Understanding the topic

Core concepts:

• Self-refine: same LLM critiques and improves.
• Critic agent: separate LLM with a stricter rubric.
• Stop when score > threshold or max iterations hit.
• Reflection works best with explicit rubrics.
• Pair with retrieval — let the critic check facts.

Syntax reference

Visual workflow / architecture:

Plan ──► Execute ──► Critique
   ▲                       │
   │                       ▼
   └──── Revise ◄──── Score < threshold?

Real-world use

Devin reflects on test failures; Cursor's agents critique their diffs; many code-gen agents auto-run tests and reflect on errors.

Best practices

• Define rubric criteria explicitly (correctness, style, length).
• Cap iterations or you'll burn dollars.

Common mistakes

• Infinite reflection loops on impossible rubrics.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is reflection in agents?
• Q2. Self-refine vs critic-agent — when each?
• Q3. Scenario: reflection hurt accuracy. Why might that happen?