LangGraph Agent Engineering

Goal: Build complex, multi-step AI workflows that are reliable, debuggable, and capable of long-running operations.

1. Core Concepts (The Graph)

• State: A explicitly defined schema (TypedDict/Pydantic) that tracks the agent’s memory snapshot.
• Nodes: Functions that perform work (call LLM, run tool, modify state).
• Edges: Logic that routes flow between nodes (Conditional edges based on LLM output).

2. Architecture Patterns

A. The ReAct Agent (Standard)

• Nodes: agent (LLM decides) <-> tools (Execute action).
• Edge: If tool call -> go to tool; If final answer -> END.

B. Plan-and-Execute (Advanced)

• Nodes: planner (Generate list) -> executor (Loop through list) -> re-planner (Update list).
• Benefit: Better for complex tasks requiring long-term reasoning.

C. Human-in-the-Loop

• Breakpoint: Insert interrupt_before=["tool_node"] to pause execution.
• Approval: Human reviews state/tool call -> Approve/Reject/Edit -> Resume graph.

3. Persistence & Memory

• Checkpointers: Use MemorySaver (for dev) or PostgresSaver (prod) to persist thread state.
• Thread ID: Always pass thread_id to graph.invoke to maintain conversation history.

4. Best Practices

• Typed State: ALWAYS define rigid TypeScript/Python interfaces for State. Do not use random dicts.
• Small Nodes: Keep nodes focused. One distinct action per node.
• Streaming: Use .stream() events to show immediate progress (tokens, node switching) to UI.

5. Example Structure (Python)

from langgraph.graph import StateGraph, END
from typing import TypedDict

class AgentState(TypedDict):
    messages: list[str]
    context: dict

def call_model(state):
    # logic...
    return {"messages": [response]}

workflow = StateGraph(AgentState)
workflow.add_node("agent", call_model)
workflow.set_entry_point("agent")
workflow.add_edge("agent", END)
app = workflow.compile()

V1.0 Migration Note:

• create_react_agent prebuilt is good for simple starts.
• For custom flows, build StateGraph manually.