LangGraph (Python)

Low-level orchestration for stateful, long-running LLM agents. LangGraph models your agent as a graph of nodes and edges with explicit state, durable checkpoints, and first-class human-in-the-loop.

Install

pip install langgraph langchain-anthropic

Version

v1.1.10 · April 28, 2026 · Python 3.10 – 3.14

Best for

Cycles, conditional routing, multi-agent coordination, HITL workflows.

Where to start

New to LangGraph? Follow the Zero → Hero path below in order (~3–4 hours). Already building? Use Jump to a topic to land on the chapter you need.

Zero → Hero

A chaptered tutorial that takes you from a first graph to a production-ready agent. Each chapter has its own page with Next → pagination.

1 · Setup & core concepts Install, state, nodes, edges, compilation · 15 min

2 · Your first agent Linear chat, conditional routing, looping, streaming · 20 min

3 · Multi-agent systems Supervisor, parallel workers, handoff · 25 min

4 · Tools ToolNode, custom executors, conditional tool use · 20 min

5 · Memory & persistence Checkpointers, Store, cross-thread memory · 25 min

6 · Streaming & debugging Stream modes, type-safe v2 API, visualization · 15 min

7 · Human-in-the-loop Interrupts, multi-stage approval, interactive debug · 25 min

8 · Middleware langchain.agents.middleware — PII, summarization, retries, HITL, cost limits · 20 min

9 · Advanced patterns ReAct, Tree-of-Thoughts, reflection, structured output, caching, Functional API · 25 min

10 · Production & troubleshooting Docker, CLI, SDK, common pitfalls · 20 min

Looking for a reference?

The Zero → Hero chapters are task-oriented tutorials. If you already know the shape of LangGraph and want exact signatures, feature tables, and breaking-change history for a single primitive, jump to the Reference section near the bottom of this page — one page per API (StateGraph, Checkpointers, Store, Functional API, Command/Send, streaming modes). Each is verified line-by-line against the installed langgraph==1.1.10 source.

Jump to a topic

Need one specific concept? Land directly on the chapter that owns it.

State graph

Conditional routing

Supervisor pattern

Parallel workers

ToolNode

Checkpointers

Store (long-term memory)

Type-safe v2 API

Interrupts (HITL)

Middleware (create_agent)

ReAct pattern

Functional API

Command / Send

Stream modes (v1 / v2)

StateGraph API

Checkpointers matrix

Store (long-term memory)

Reference

Every page below is introspected against the installed langgraph==1.1.10 / langgraph-checkpoint==4.0.3 / langgraph-checkpoint-sqlite==3.0.3 / langgraph-checkpoint-postgres==3.0.5 source. If a signature disagrees with an older blog post or docs page, the source wins.

StateGraph API reference Constructor, add_node (all 4 overloads), add_edge, add_conditional_edges, add_sequence, compile, Runtime, retry/cache/defer/destinations.

Checkpointers reference Feature matrix for InMemorySaver, Sqlite, AsyncSqlite, Postgres, AsyncPostgres, ShallowPostgres — plus durability modes and connection strings.

Store (long-term memory) reference BaseStore, InMemoryStore, namespaces, IndexConfig (vector search), TTLConfig, InjectedStore for tools.

Functional API reference @entrypoint, @task, entrypoint.final, previous, injecting Runtime, feature matrix vs StateGraph.

Command / Send reference Command(update, goto, resume, graph), Send for fan-out, interrupt + resume, Command.PARENT, Overwrite.

Streaming modes reference All seven stream_mode values, v1 vs v2 typing, StreamPart discriminated union, GraphOutput, subgraphs=True.

Single-page comprehensive guide All chapters in one long page — best for search and printing.

Recipes RAG, support routing, research agent, doc pipelines, long-term memory conversations.

Diagrams 20+ Mermaid architecture diagrams.

Production guide Deployment, scaling, monitoring, disaster recovery.

Advanced error recovery

Observability

Performance optimization

Streaming server (FastAPI)

What’s new in v1.1.x (April 2026)

• Type-safe streaming (v2 API) — opt-in with version="v2" on astream(); yields StreamPart objects with type, ns, data fields (import from langgraph.types).
• Type-safe invoke (v2 API) — ainvoke(..., version="v2") returns a GraphOutput (from langgraph.types) with .value (final state dict) and .interrupts (tuple of interrupt points).
• Pydantic/dataclass coercion — invoke() auto-coerces input dicts to the declared state type.
• Python 3.10 – 3.14 support — Python 3.9 dropped in the 1.1 line.
• Time-travel fixes — replays no longer reuse stale RESUME values with interrupts and subgraphs.

Verified against installed library

Introspected against langgraph==1.1.10, langgraph-checkpoint==4.0.3, langgraph-prebuilt==1.0.10, langgraph-checkpoint-sqlite==3.0.3, langgraph-checkpoint-postgres==3.0.5 in April 2026.

Known corrections vs older drafts / third-party guides:

• langgraph.llm_hooks and langgraph.cache.cache_node do not exist. The middleware system is langchain.agents.create_agent(middleware=[...]); see Chapter 8.
• create_react_agent in langgraph.prebuilt is deprecated in v1.0 in favor of langchain.agents.create_agent. It still runs, with a warning.
• AgentState / AgentStatePydantic moved from langgraph.prebuilt to langchain.agents.
• SqliteSaver.from_conn_string(...) and PostgresSaver.from_conn_string(...) are context managers (with ... as saver:), not factories that return a saver directly. Some earlier examples omitted the with; see the Checkpointers reference.
• Interrupt.ns, Interrupt.when, Interrupt.resumable were removed in v0.6; only Interrupt.value and Interrupt.id remain.
• checkpoint_during=False on invoke/stream is deprecated — use durability="exit" (or "sync" / "async").
• ShallowPostgresSaver is deprecated as of langgraph-checkpoint-postgres 2.0.20. Use PostgresSaver + durability="exit".

Install, minimal example

from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import InMemorySaver
from typing_extensions import TypedDict

class State(TypedDict):
    message: str
    response: str

def process(state: State) -> dict:
    return {"response": f"Processed: {state['message']}"}

builder = StateGraph(State)
builder.add_node("process", process)
builder.add_edge(START, "process")
builder.add_edge("process", END)

graph = builder.compile(checkpointer=InMemorySaver())

result = graph.invoke(
    {"message": "Hello"},
    config={"configurable": {"thread_id": "user-1"}},
)
print(result)

Ready for the full walk-through? Start Chapter 1 →

---

Revision history

Date	Version	Changes
2026-04-28	1.1.10	Patch release; version card and reference pages updated 1.1.9 → 1.1.10; langgraph-checkpoint updated 4.0.2 → 4.0.3. All key symbols verified against installed 1.1.10 (.routine-envs/main-py-0428).
2026-04-22	1.1.9	Added six source-verified reference pages: StateGraph, Checkpointers, Store, Functional API, Command / Send, Streaming modes. Corrected stale guidance: create_react_agent deprecated in v1.0; from_conn_string is a context manager; ShallowPostgresSaver deprecated.
2026-04-21	1.1.8	Index redesigned into Zero → Hero chaptered tutorial + Jump-to-topic grid; duplicated sections removed.
2026-04-20	1.1.8	Version pin updated to 1.1.8; frameworks.ts metadata aligned.
April 16, 2026	1.1.6	Python 3.9 dropped; type-safe v2 streaming/invoke API; Pydantic/dataclass coercion; Python 3.14 support; time-travel fixes.
November 2025	1.0.3	Node caching; deferred nodes; pre/post model hooks; cross-thread memory; Python 3.13 support.