LangGraph
Low-level orchestration framework for stateful, long-running LLM agents, with checkpointing, human-in-the-loop pauses and a managed deployment tier through LangSmith.
Overview
LangGraph is the orchestration layer beneath LangChain, and the two are worth separating in your head. LangChain gives you a fast path to a working agent. LangGraph gives you a directed graph where each node is a function, edges decide what runs next and everything shared lives in a typed state object. A checkpointer writes that state after every step, so a run can pause for a human sign-off on Tuesday and continue on Thursday from exactly where it stopped. Both reached v1.0 in October 2025 under an LTS compatibility promise, and langgraph sits at 1.2.11 on PyPI as of August 2026.
The teams it fits are the ones with agents that branch, retry, call other agents and wait on people. Approval gates in a procurement flow. A support agent that escalates and sits idle for two days. A long-running research job that must survive a pod restart. Time travel, rewinding to an earlier checkpoint and branching from it, turns out to be the debugging tool people reach for once a system is live, and it is close to impossible to retrofit onto a framework that keeps state in memory.
Skip it if your agent is one model, five tools and a loop. The setup cost will not pay back and the OpenAI Agents SDK or Pydantic AI will ship the same thing sooner. Budget carefully before committing to the managed side too: the framework is MIT and free forever, but LangSmith at $39 a seat plus metered compute units is a different conversation with finance than a flat platform fee, and the metering model has already changed once in the past year.
Key Features
- ✓ Graph execution model with explicit nodes, edges and a typed shared state object, so agent control flow is code you can read rather than prompt behaviour you hope for
- ✓ Checkpointers that persist state to Postgres or SQLite, letting a run pause for days and resume mid-graph on a different machine
- ✓ Interrupt and resume primitives for human approval steps, plus time travel to rewind a run to an earlier checkpoint and branch from there
- ✓ On the 1.2.x line (1.2.11 on PyPI as of August 2026) under the 1.0 LTS backward-compatibility promise made in October 2025
- ✓ LangSmith tracing wired in by default, covering per-node latency, token counts and eval runs
- ✓ Deployment paths from self-hosted containers to LangSmith Serverless, with sandboxes and an LLM gateway added to the platform during 2026
Where it holds
- • The state machine is explicit. When an agent loops forever or silently drops a tool result, you can point at the node that did it.
- • Durable checkpointing is genuinely production grade, and the Postgres saver is boring in the right way.
- • Adoption is deep enough that hiring and debugging get easier. Klarna, Uber, LinkedIn, Cisco and Nvidia sit on the public customer wall.
- • MIT licence with no feature gate on the framework itself, so self-hosting is a real option and not a crippled tier.
Where it breaks
- • The learning curve is real. State, reducers, edges and checkpointers are four concepts to absorb before you write anything useful, and a team building a linear tool-calling agent will find Pydantic AI does the job with a fraction of the ceremony.
- • It inherits LangChain's reputation for leaky abstractions, and a good number of the surrounding ecosystem integrations are still thin wrappers of uneven quality.
- • Managed pricing moved to Compute Units and Storage Units, which makes forecasting a monthly bill harder than the flat per-node rate did.
- • Observability pulls you toward LangSmith by default. Routing traces to something you already operate is possible but it is not the paved road.
My Take
State is the real problem, not orchestration, and that is the thing this answers well. The graph is explicit, checkpoints land in Postgres and a run that dies on a tool call resumes at the node it died on rather than the top of the loop. The cost is concept overhead. For a straightforward tool-calling agent, Pydantic AI gets you there with far less machinery. Version 1.2.11 is stable and the 1.0 LTS compatibility promise from October 2025 has held, which matters more than any benchmark for something you intend to run for two years.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- MIT
- Starting at:
- MIT-licensed and free to self-host. The managed side is LangSmith: Developer at $0 for one seat with 5k base traces a month, Plus at $39 per seat per month with 10k base traces and one free Serverless (Small) deployment, Enterprise custom. Extra deployments meter as LangChain Compute Units at $1.50 per LCU and Storage Units at $1.00 per LSU. Note that the older $0.001-per-node-execution and standby-minute rates that most third-party pricing pages still quote no longer appear on langchain.com/pricing, checked August 2026.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where LangGraph holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to LangGraph
OpenAI Agents SDK
open-sourceMinimal open-source Python and TypeScript framework for building agents with handoffs, guardrails, sessions and sandboxed code execution, usable with non-OpenAI models.
AgentScope
open-sourceApache 2.0 agent framework from Alibaba Tongyi Lab's SysML team. Version 2.0 folded the separate Runtime project into the core, so sandboxing, agent serving and observability now ship as one...
Temporal
freemiumDurable execution platform that keeps long-running agent and backend workflows alive across crashes, restarts and outages, with SDKs in Python, Go, Java, TypeScript and .NET.
LlamaIndex
freemiumOpen-source Python and TypeScript framework for RAG and agentic workflows, paired with LlamaCloud for document parsing, extraction and managed indexes.