Skip to main content

LangGraph

LangGraph is a graph-based orchestration framework for building stateful LLM applications. The key idea is to replace “prompt spaghetti” with a state machine you can debug, test, replay, and evolve.

Learning goals​

  • Model an agent as a graph of nodes (steps) with explicit transitions
  • Add guard conditions + retries without fragile prompt hacks
  • Make runs reproducible (state snapshots)

The mental model​

A LangGraph app is:

  • a State object (your single source of truth)
  • a set of Nodes (pure-ish functions that read/write state)
  • Edges (transitions) that decide what runs next
text
[User Input]
   ↓
[Planner] → [Tool Call] → [Synthesis]
   ↘︎                ↘︎
    └──────[Retry/Repair]───┘

Why graphs beat chains​

  • Branching: handle “if tool fails, retry with fallback” cleanly
  • Observability: log node-by-node decisions
  • Human in the loop: insert approvals / checkpoints
  • Testing: unit test nodes without a live model

Minimal pseudo-code​

python
# Pseudo-code (conceptual)

class State(dict):
    # keys: question, plan, tool_results, answer, errors
    ...

def planner(state: State) -> State:
    state["plan"] = "..."
    return state

def tool_node(state: State) -> State:
    state["tool_results"] = {"search": "..."}
    return state

def synthesizer(state: State) -> State:
    state["answer"] = "..."
    return state

# graph.add_node("planner", planner)
# graph.add_edge("planner", "tool")
# graph.add_edge("tool", "synth")

Practical patterns to copy​

  • Router node: decides which tool to call next based on state.
  • Repair loop: if structured output validation fails, branch to a repair node.
  • Budget node: stops the run when tokens/steps exceed a threshold.

Mini-lab (optional)​

Build a “research agent graph”:

  • planner → web search tool → read tool → synthesize
  • add a retry branch if citations are missing
  • log every node output to a JSONL file

Where this fits​

LangGraph is the orchestration layer. Your tools can be local function calls, or exposed via MCP Server.