LangGraph
A developer framework for long-running, stateful agents.
Group: Build-your-own platforms and infrastructure
LangChain documents LangGraph as a low-level runtime for long-running stateful agents, including durable execution and human-in-the-loop support.
The decision in plain English
A framework is a building block, not a ready-made assistant; compare developer requirements and runtime choices.
Best fit
Developers implementing stateful agent workflows.
Poor fit
A finished assistant that works without development.
Documented capabilities
Vendor statements, not verified task outcomes.
Permissions and human control
- Developers place dynamic interrupts before side effects and implement approval/rejection handling. Framework support alone does not enforce a universal permission policy.
Documented controls do not prove that an agent always follows them.
Limits and open questions
- A durable checkpointer and stable thread ID are prerequisites for resumable production workflows.
- Resuming restarts the interrupted node, so pre-interrupt side effects can repeat.
- Model, database, logging and hosting costs and data terms are separate from the framework.
Privacy and security
Workflow state can be persisted by the chosen checkpointer. Retention, access, deletion, model training and logs depend on the developer-selected services and configuration; no framework-wide privacy guarantee or retention period is established.
Feature-by-feature: what is documented and what is not
Official sources checked 22 October 2026. A documented feature is a vendor statement, not proof of success.
| Feature | Evidence status | Detail and limit |
|---|---|---|
| Product type | Developer framework | LangGraph is a low-level runtime for stateful agent workflows, not a finished assistant with ready-made accounts or permissions. Developers choose models, tools and deployment. Source |
| Checkpoint storage | Configure persistence | Compile the graph with a checkpointer. In-memory examples are not durable production storage; SQLite, Postgres and other adapters have separate installation and operating needs. Source |
| Pause for input | Implementation prerequisites | Dynamic interrupt() needs a checkpointer, a thread ID and a JSON-serializable payload. It saves state and waits until the application resumes it; the developer decides where to pause. Source |
| Resume identity | Keep the same thread | Command(resume=...) supplies the response to the interrupt. Reuse the thread ID to load the checkpoint; a new ID starts an empty thread rather than continuing the old one. Source |
| Replay and side effects | Node code can run again | Resuming restarts the interrupted node from its beginning. Operations before interrupt can repeat, so use idempotent operations or separate side-effect nodes and deduplication. Source |
| Human approval | Not automatic for every tool | Developers must put the interrupt before the actual action and handle approve, edit and reject responses. Static breakpoints are for debugging, not the recommended human-in-the-loop pattern. Source |
| Licence and cost | MIT is not free operation | The repository lists an MIT licence. Model inference, hosting, storage and optional LangSmith services remain separate; no complete workload price is verified. Source |
| Privacy and retention | Your deployment decides | Checkpointing persists workflow state; it is not a promise of deletion or model-training opt-out. Review the chosen store, model provider, logging and hosting terms. No framework-wide retention period is established. Source |
How to start with LangGraph narrowly
- Build a small graph with a single read-only tool and explicit success/failure states.
- Choose a checkpointer and stable thread ID. Use a durable store for production rather than an in-memory demo.
- Place interrupt() before a side effect; implement approve, edit and reject responses with Command(resume=...).
- Check replay behavior and deduplicate writes before connecting real accounts. Review model, hosting and storage costs and data terms.
Official product/setup source. These are documentation-based steps, not a hands-on setup test.
Common questions about LangGraph
What is LangGraph for?
Developers implementing stateful agent workflows. The vendor describes it as: LangChain documents LangGraph as a low-level runtime for long-running stateful agents, including durable execution and human-in-the-loop support.
Who should look elsewhere?
A finished assistant that works without development.
How much does LangGraph cost?
LangGraph is an open-source framework (MIT licence, per the GitHub repository, re-verified 22 October 2026). This is not a promise of free agent operation: model inference, hosting and optional LangSmith services have separate costs. No complete deployed-workload price is verified.
What can it access and what needs my approval?
Developers place dynamic interrupts before side effects and implement approval/rejection handling. Framework support alone does not enforce a universal permission policy.
What happens to my data?
Workflow state can be persisted by the chosen checkpointer. Retention, access, deletion, model training and logs depend on the developer-selected services and configuration; no framework-wide privacy guarantee or retention period is established.
Has AgentComparison tested it?
No. This record is built from the vendor's own documentation, checked 2026-10-22. Vendor claims are attributed, not confirmed.
How to evaluate LangGraph yourself
- Pick one low-risk task that matches its documented capabilities: durable execution and persistence.
- Write the expected output and what counts as failure before you start.
- Connect the minimum access needed, and check approval settings first.
- Run it twice and compare, then review every action it took.
- Check cost after any free allowance. Unknown price is not free.
Related
Documented use cases: Building your own agent.
- Microsoft Copilot Studio (platform): Teams building business-data agents for several channels.
- Case (product): Developers who want an agent to keep logged-in sessions and files between runs, for jobs such as scheduled browser tasks, and who are comfortable running or paying for the machine and supplying their own model key.
- Build with LangGraph or use a ready-made agent?
- AI agent use cases by job
- Privacy and approval checklist
- Compare two agents side by side
- Latest agent updates
Sources and history
- https://docs.langchain.com/oss/python/langgraph/overview · official · read 22 October 2026
- https://docs.langchain.com/oss/python/langgraph/checkpointers · official · read 22 October 2026
- https://docs.langchain.com/oss/python/langgraph/interrupts · official · read 22 October 2026
- https://github.com/langchain-ai/langgraph · official · read 22 October 2026
Change history
- 2026-10-22: Re-verified vendor documentation claims (durable execution, human-in-the-loop interrupts, persistence, memory) against current LangGraph docs; MIT licence and active maintenance confirmed via the GitHub repository API. No content corrections needed; no hands-on testing.
- 2026-10-14: Expanded product-specific runtime, approval, access and privacy evidence with narrow setup steps. Documentation research, not hands-on testing.
- 2026-10-01: Initial documentation research. No hands-on testing.
Documentation overview expanded with current pricing/access sources. No hands-on results claimed.
Report a correction