---

name: LangGraph description: Workflow patterns and gotchas for LangGraph. Directs to RAG for implementation.

LangGraph Workflow

When to Choose LangGraph

• Need stateful, cyclic agent workflows
• Building complex multi-step reasoning
• Want conditional branching and loops
• Require human-in-the-loop checkpoints

Decision Framework

Graph Pattern Selection

Query RAG: mcp__agentic-rag__query_sdk("pattern example", sdk="langgraph", mode="build")

Critical Gotchas

These cause debugging nightmares:

1. State must be TypedDict - Not a regular dict; needs type annotations
2. Node returns partial state - Return only keys you're updating, not full state
3. END is special - Import it: from langgraph.graph import END
4. Edges define flow - Forgetting an edge = node never reached
5. Conditional edges return node names - Return the string name, not the function
6. Compile before run - graph.compile() is required before invoke
7. Checkpointer for memory - Without it, state resets each run
8. State channels merge - Multiple updates to same key need reducer

Workflow: Building a LangGraph Agent

Step 1: State Definition

RAG Query: mcp__agentic-rag__query_sdk("TypedDict state definition", sdk="langgraph", mode="build")

Define your state schema with TypedDict and Annotated for reducers.

Step 2: Node Functions

RAG Query: mcp__agentic-rag__query_sdk("graph node function", sdk="langgraph", mode="build")

Each node takes state, returns partial state update.

Step 3: Graph Construction

RAG Query: mcp__agentic-rag__query_sdk("StateGraph add_node add_edge", sdk="langgraph", mode="build")

Step 4: Edge Definition

RAG Query: mcp__agentic-rag__query_sdk("conditional_edges routing", sdk="langgraph", mode="build")

Step 5: Compilation

RAG Query: mcp__agentic-rag__query_sdk("graph compile checkpointer", sdk="langgraph", mode="build")

Step 6: Execution

RAG Query: mcp__agentic-rag__query_sdk("compiled graph invoke stream", sdk="langgraph", mode="build")

Common Error Patterns

Graph Patterns

Decision Tree

Nodes for each decision point, conditional edges for branching. RAG Query: mcp__agentic-rag__query_sdk("decision tree graph", sdk="langgraph", mode="build")

ReAct Agent

Reason-Act-Observe loop with tool calling. RAG Query: mcp__agentic-rag__query_sdk("react agent pattern", sdk="langgraph", mode="build")

Plan-and-Execute

Planning node, execution loop, verification. RAG Query: mcp__agentic-rag__query_sdk("plan execute pattern", sdk="langgraph", mode="build")

Advanced Features

Query RAG when you need:

• Streaming: "langgraph streaming events"
• Subgraphs: "nested subgraph composition"
• Human-in-loop: "interrupt checkpoint approval"
• Time travel: "state history replay"
• Parallel branches: "parallel node execution"