SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillslanggraph-workflow

langgraph-workflow

Agents & models

View repository

Updated 10-15 19:54

About this skill

langgraph-workflow — Stateful Multi-Agent Graph Design

LangGraph is the industry-standard framework for production multi-agent workflows. Build StateGraph-based pipelines with checkpointing, conditional edges, HITL gates, and parallel execution.

When to use this skill

  • Building a multi-step AI pipeline where intermediate state must persist between nodes
  • Implementing retry/loop logic (e.g., Editor re-routes to upstream node on low quality)
  • Adding Human-in-the-Loop approval gates before destructive operations
  • Running nodes in parallel to reduce latency
  • Recovering from partial failures via checkpoint resume
  • Integrating LangSmith tracing for observability (@traceable)

Core Concepts

StateGraph and State Schema

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

class AgentState(TypedDict):
    messages: list[dict]
    draft: str
    quality_score: float
    retry_count: int
    error: str | None

graph = StateGraph(AgentState)

Node functions

Each node is an async function that receives state and returns a partial update:

async def researcher_node(state: AgentState) -> AgentState:
    result = await llm.ainvoke(state["messages"])
    return {"draft": result.content}

async def editor_node(state: AgentState) -> AgentState:
    score = evaluate_quality(state["draft"])
    return {"quality_score": score}

graph.add_node("researcher", researcher_node)
graph.add_node("editor", editor_node)

Edges and conditional routing

def route_after_editor(state: AgentState) -> str:
    if state["quality_score"] < 0.7 and state["retry_count"] < 3:
        return "retry"
    return "done"

graph.set_entry_point("researcher")
graph.add_edge("researcher", "editor")
graph.add_conditional_edges(
    "editor",
    route_after_editor,
    {"retry": "researcher", "done": END},
)

compiled = graph.compile()

Parallel node execution

# Run two nodes in parallel with Send API
from langgraph.types import Send

def fan_out(state: AgentState) -> list[Send]:
    return [
        Send("analyzer_a", {"input": state["draft"]}),
        Send("analyzer_b", {"input": state["draft"]}),
    ]

graph.add_conditional_edges("splitter", fan_out)

Instructions

Step 1: Define state schema

Use TypedDict for all state fields. Include error handling fields:

class WorkflowState(TypedDict):
    task: str
    results: dict[str, str]    # keyed by node name
    quality: float
    retry_count: int
    metadata: dict             # token usage, timing, etc.
    error: str | None

Step 2: Implement nodes

One node per logical step. Keep nodes pure (no side effects outside state):

async def planning_node(state: WorkflowState) -> WorkflowState:
    plan = await llm.ainvoke(f"Plan this task: {state['task']}")
    return {"results": {**state["results"], "plan": plan.content}}

Step 3: Add checkpointing (for HITL or resumable workflows)

from langgraph.checkpoint.memory import MemorySaver

checkpointer = MemorySaver()
compiled = graph.compile(checkpointer=checkpointer)

# Invoke with thread_id for persistence
result = await compiled.ainvoke(
    initial_state,
    config={"configurable": {"thread_id": "session-abc"}},
)

Step 4: Add HITL interrupt gates

# Interrupt BEFORE a node runs
compiled = graph.compile(
    checkpointer=checkpointer,
    interrupt_before=["deploy_node"],  # pause before deploy
)

# Resume after human approval:
result = await compiled.ainvoke(
    None,  # None = resume from checkpoint
    config={"configurable": {"thread_id": "session-abc"}},
)

Step 5: Add LangSmith tracing

import os
os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_API_KEY"] = "your-key"

# Nodes auto-traced. Add @traceable for custom spans:
from langsmith import traceable

@traceable(name="custom-eval")
def evaluate_quality(text: str) -> float:
    ...

Step 6: Stream node events

async for event in compiled.astream(initial_state):
    node_name = list(event.keys())[0]
    node_output = event[node_name]
    print(f"[{node_name}] completed")

Examples

Example 1: Quality-gated writing pipeline

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

class WriterState(TypedDict):
    topic: str
    draft: str
    quality: float
    retries: int

async def write_node(state: WriterState) -> WriterState:
    draft = await llm.ainvoke(f"Write about: {state['topic']}")
    return {"draft": draft.content}

async def review_node(state: WriterState) -> WriterState:
    score = await score_quality(state["draft"])
    return {"quality": score}

def route(state: WriterState) -> str:
    if state["quality"] < 0.7 and state["retries"] < 3:
        return "rewrite"
    return "done"

graph = StateGraph(WriterState)
graph.add_node("write", write_node)
graph.add_node("review", review_node)
graph.set_entry_point("write")
graph.add_edge("write", "review")
graph.add_conditional_edges("review", route, {"rewrite": "write", "done": END})
app = graph.compile()

Example 2: Multi-agent research with parallel nodes

async def fetch_sources(state):
    return [
        Send("fetch_web", {"query": state["query"]}),
        Send("fetch_docs", {"query": state["query"]}),
    ]

graph.add_conditional_edges("planner", fetch_sources)
graph.add_edge("fetch_web", "synthesizer")
graph.add_edge("fetch_docs", "synthesizer")

Example 3: HITL deployment gate

compiled = graph.compile(
    checkpointer=MemorySaver(),
    interrupt_before=["deploy"],
)
# Agent pauses at "deploy". Human reviews, then:
await compiled.ainvoke(None, config={"configurable": {"thread_id": tid}})

Best practices

  1. Type all state fields — use TypedDict; avoid dict[str, Any] at top level
  2. Return partial updates — nodes return only changed fields, not the full state
  3. Cap retry counts in state — always include retry_count: int and check it in routing
  4. Use checkpointing for any HITL — MemorySaver() for development, SqliteSaver for production
  5. Stream in production — astream() prevents timeout on long-running graphs
  6. Trace with LangSmith — set LANGCHAIN_TRACING_V2=true before going to production

References