11. Reflection Agent¶

Overview¶

The Reflection Agent implements a powerful AI pattern where an LLM critiques its own output and iteratively improves it through a generate → reflect → revise loop. This section builds a complete reflection agent using LangGraph that takes a Twitter post and refines it through multiple rounds of self-critique.

Architecture at a Glance¶

flowchart TD
    START((▶ START)) --> GEN["✍️ Generate\n(Create/Revise Tweet)"]
    GEN -->|"messages < 6"| REF["🔍 Reflect\n(Critique Tweet)"]
    GEN -->|"messages ≥ 6"| END((⏹ END))
    REF --> GEN

    style START fill:#10b981,color:#fff
    style END fill:#ef4444,color:#fff
    style GEN fill:#4a9eff,color:#fff
    style REF fill:#f59e0b,color:#fff

Lesson Map¶

#	Lesson	Focus
1	What Are We Building?	The reflection pattern — why self-critique produces better outputs
2	Project Setup	Poetry, dependencies, environment variables, sanity checks
3	Creating the Chains	The generation and reflection chains using LCEL
4	Defining the LangGraph	State, nodes, edges, conditional routing, and graph compilation
5	LangSmith Tracing	Executing the graph, tracing the execution, and analyzing results

Key Technologies¶

Technology	Role
LangGraph	Graph-based workflow orchestration (generate ↔ reflect loop)
LangChain	Prompt templates, LCEL chains, message types
OpenAI GPT-3.5	LLM for both generation and reflection
LangSmith	Observability — traces every node execution and LLM call
Poetry	Dependency management and virtual environment