LangGraph: our review

LangGraph is the AI agent orchestration framework built by LangChain, the vendor behind the Python library of the same name. Where LangChain provides the basic building blocks for calling a model, LangGraph provides the architecture for making several calls work together: planning a sequence, remembering intermediate states, and asking a human before crossing a critical step.

The framework is open source under the MIT licence and installs with one command. The paid LangSmith cloud platform adds managed deployment, tracing, evaluation and code-execution sandboxes, but it isn't required: a LangGraph agent runs perfectly well on your own server.

What it actually does

A LangGraph agent is modelled as a directed graph: nodes (each reasoning or action step) linked by conditional edges. The agent starts from an initial state — a user's question, for example — and moves through the graph, deciding at each node which branch to take.

Two properties set LangGraph apart from competing frameworks. The first is persistent memory: the agent keeps the full history of its decisions and can return to it, which allows long chains of reasoning without losing the thread. The second is native Human-in-the-Loop: at any node in the graph, you can pause execution and wait for human approval before continuing. For an agent that drafts a contract, approves an order or triggers a payment, this point is decisive.

In practice, an SME deploying an AI agent to automate handling customer requests can build a graph that reads the incoming message, classifies it, searches the knowledge base for an answer, drafts a reply, and submits it to an advisor for approval before sending.

Why it interests us

LangGraph is the framework we use most often for bespoke AI agent development projects. The reason comes down to one word: reliability. Most agent frameworks work well on simple cases but become unpredictable as soon as a task has more than three or four steps. LangGraph's graph-based model, paired with memory and human control, gives a framework for building agents whose behaviour can be predicted and audited — the minimum condition for professional use.

The fact that it can be self-hosted is also critical for our clients in regulated sectors: a law firm or an accounting practice can run the agent on its own server and guarantee to its clients that case files never leave the practice.

Where it falls short

The first obstacle is the technical requirement. LangGraph has no graphical interface: everything runs through Python or JavaScript code. An SME without a developer can't pick it up on its own, and even with a developer the learning curve is real. It's a control framework, not an immediate-productivity tool.

The LangSmith cloud platform, which brings managed deployment, tracing and evaluation, is hosted in the United States with no option for European residency. For an organisation that wants both the simplicity of the cloud and data sovereignty, it doesn't add up.

Finally, LangChain's privacy policy doesn't clarify how data is used for training. The "research and development" purpose potentially covers this use, and no explicit exclusion is made. For a buyer weighing the legal risk of passing business data through LangSmith, this lack of an answer is a sticking point.

Our recommendation

If you have a developer on the team and a complex business process to automate with a reliable, auditable agent, LangGraph is the best open-source framework available. Self-hosted, it meets even the strictest confidentiality requirements.

If you're looking for a no-code tool to prototype an agent quickly, or want an EU-hosted cloud platform, this isn't the right choice: look instead at n8n or European managed solutions such as Scaleway.