LangChain vs LangGraph: what actually differs, and when to use which

LangChain and LangGraph come from the same team, share an ecosystem, and are constantly compared — usually by someone deciding what to build an agent on this quarter. The confusion is understandable: one grew out of the other, the documentation cross-links heavily, and both can technically build an “agent”.

Here is the practical difference, when each fits, and the part most comparisons skip: the work that remains after you pick one.

LangChain vs LangGraph: the short answer

LangChain is a component library: hundreds of integrations for models, vector stores, retrievers, and tools, plus composable chains for sequential LLM pipelines (prompt → model → parser → next step). It is at its best when data flows one way.

LangGraph is an orchestration framework for stateful agents: you define nodes and edges in a graph, and the framework manages state, branching, loops, persistence via checkpointers, and human-in-the-loop interrupts. It exists because real agents do not flow one way — they loop, branch, retry, and wait for people.

Rule of thumb: pipelines → LangChain; agents → LangGraph. They compose — LangGraph nodes routinely call LangChain components.

What each is best at

Dimension	LangChain	LangGraph
Mental model	Chains — directed pipelines	Graphs — nodes, edges, cycles, state
State	Mostly stateless between steps	First-class state object, checkpointed
Loops & branching	Awkward — chains want to move forward	Native — cycles and conditional edges
Human-in-the-loop	Middleware-based since v1 (HumanInTheLoopMiddleware)	Interrupts and resumable runs at the graph level
Multi-agent	Assembled by hand from agents and middleware	Supervisor, swarm, and subgraph patterns supported
Best for	RAG pipelines, extraction, simple assistants	Agents that decide, retry, and coordinate
Learning curve	Gentler start, abstractions run deep	Steeper start, more explicit control

When to use LangChain

Choose plain LangChain when the shape of your problem is a pipeline: retrieval-augmented answering, document extraction, summarization, classification. You get the integration catalog — genuinely the broadest in the ecosystem — without graph machinery you do not need. Be aware of the trade-off teams cite most: deep abstraction layers and a brisk pace of breaking changes that your codebase inherits.

When to use LangGraph

Choose LangGraph the moment your system needs to make decisions about its own flow: tool-using agents that loop until done, workflows that pause for human approval, supervisors delegating to specialists. Explicit graphs are easier to reason about than prompt-buried control flow, and checkpointed state gives you resumability. The cost: you are now maintaining orchestration code, and the service layer around it is still yours to build or to buy from the vendor.

What the framework alone does not give you — and what you buy to fill it

Credit where it is due: LangGraph OSS already ships the hard runtime internals — a first-class state object, checkpointed persistence, streaming primitives, and durable execution. The library is not the gap. The gap is the service around it, and with LangChain that service is assembled from several vendor surfaces — some free, some paid — rather than shipped with the library:

• An API server with streaming your product can consume — the LangGraph server; langgraph dev runs it locally on a free LangSmith key, while the standalone self-hosted deployment needs a LangSmith licence key and the managed control plane and HA sit in paid LangSmith Deployment tiers
• Authentication, token scopes, and tenant isolation — custom auth handlers you write against that server, plus whatever tenant scoping your product needs
• An admin surface non-engineers can operate — Studio is an agent IDE for developers (free against a local server), so the non-engineer surface is Fleet, a module inside LangSmith
• Observability, evaluations, and audit — LangSmith, free on the Developer tier and paid above it
• Tool security: per-agent scoping, forwarded end-user identity, confirmation gates on the individual tool — this one you still write yourself

None of that is a criticism of the engineering; it is a statement about shape. You end up operating a framework plus three or four surfaces from the same vendor, each with its own pricing and upgrade path. SyntheticBrew takes the other shape: one managed runtime where agents are configuration, delegation and spawning have policies, tools carry your users’ real identity, and every step is streamed and audited — API, admin, sessions, and audit from the same service. Your product integrates over REST + SSE in any language, and no framework abstractions enter your codebase.

If you are weighing the frameworks because an agent feature needs to ship, read our LangChain alternatives comparison — it covers when a framework is the right call and when a runtime gets you to production faster.

LangChain vs LangGraph FAQ