LangGraph

What it is

LangChain describes LangGraph as an agent runtime and a low-level orchestration framework, one that tries to balance an operator's control over an agent with the agent's own room to act ^[1]. An "agent" here means a language model that does more than answer: it is given tools and loops through steps, deciding what to do next and acting on that decision. The "runtime" is the layer that actually runs that loop — it holds state between steps, chooses what happens next, and executes the calls the agent asks for. "Low-level" means the framework supplies building blocks rather than a fixed application structure, so the developer assembles the control flow instead of accepting a prescribed one.

How the control flow is expressed

The framework's primitives cover single-agent, multi-agent and hierarchical control flows inside one framework ^[1]. In a single-agent flow, one model with its tools handles the whole task. In a multi-agent flow, the work is split among several agents, each with its own role and tools. A hierarchical flow adds a layer above them, where one agent or component directs the others. Because all three shapes come from the same set of primitives, a project can start small and add agents later without moving to a different system ^[1].

Memory, streaming and human oversight

LangGraph ships memory as a built-in feature: it stores conversation histories and keeps context across sessions, so an agent that has dealt with a user before can continue from where it left off rather than starting from nothing ^[1].

It also provides native token-by-token streaming, so an agent's reasoning and actions appear in real time as they are produced ^[1]. This matters for anything a person is watching: the gap between a request and a finished answer can be many seconds, and streaming turns that wait into visible progress.

The third piece is human-in-the-loop checks, which can be added to steer and approve what an agent does ^[1]. Instead of letting an agent run to completion, a developer can place a point in the flow where a person reviews the next step and approves it, rejects it, or redirects it. This is the same control-versus-agency balance in a concrete form: the more autonomy an agent is given, the more the approval points carry.

The 1.0 release

LangGraph 1.0 became generally available on 23 October 2025, according to the LangChain changelog ^[2]. LangChain described the release as the first stable major release in the durable agent framework space ^[3]. Stability here means the interface developers build against is no longer expected to shift under them — a practical concern for teams that maintain agent systems over months.

The release kept full backward compatibility, and its only notable change was deprecating the langgraph.prebuilt module ^[3]. Full backward compatibility means code written against earlier versions keeps running; deprecating a module means it is being phased out, with its functionality available elsewhere in the framework rather than removed outright.

Where it sits

LangChain said that LangChain and LangGraph together had 90 million monthly downloads at the time of its Series B ^[3]. That figure measures adoption rather than quality, but it shows how widely the framework is used as a base layer for agent systems. The position LangChain gives it — low-level, with memory, streaming and approval as built-in primitives — places it below complete applications and above raw model APIs: the layer where an agent's control flow is defined.