LangGraph
Low-level orchestration framework and runtime for building stateful, long-running agents as explicit graphs. From LangChain. Roughly 35K GitHub stars as of 2026-06-18, used in the LangChain platform stack for durable agent execution.
What it is
LangGraph models an agent as a graph of nodes (functions or LLM calls) connected by edges, with a shared, persisted state object threaded through every step. Unlike higher-level "agent in three lines" wrappers, it is intentionally low-level: the developer owns control flow, branching, loops, and where the agent pauses. This makes it suited to long-running, stateful workflows rather than single-shot chat. It ships for Python (pip install -U langgraph) and JavaScript. Source: langchain-ai/langgraph README, https://github.com/langchain-ai/langgraph, 2026-05-31
Why teams pick it
- Durable execution - graphs checkpoint state, so an agent can crash, resume, or be replayed from any node without losing context.
- Human-in-the-loop - execution can pause at a node, surface state for approval or edits, then continue.
- Explicit state - the shared state schema is first-class, which makes complex multi-agent topologies (supervisors, sub-graphs, cycles) auditable.
- Memory - short-term (within a thread) and long-term (across threads) memory are built in, and pair with external stores like Neo4j Agent Memory.
It is decoupled from the broader LangChain library, so LangGraph can be adopted on its own. LangGraph Platform is the companion managed deployment and runtime for serving these agents, with a Studio for visual debugging.
LangChain's current product map is a useful ontology check: LangChain is the agent framework, LangGraph is the orchestration runtime, Deep Agents is the higher-level agent harness built on top, and LangSmith is the tracing/evaluation/deployment layer. LangSmith Sandboxes add the environment side of that map: managed sandboxes, service URLs, mounts, and LangSmith Sandbox Auth Proxy for outbound credentials and egress policy. That cleanly matches Kevin's Agent Harness / Agent Runtime Architecture / Agent Control Plane Architecture / Agent Security Model split. Source: LangGraph docs, https://docs.langchain.com/oss/python/langgraph/overview, 2026-06-18; LangSmith sandbox docs, 2026-07-04
How it compares
LangGraph trades the brevity of the OpenAI Agents SDK (handoffs, minimal ceremony) for explicit, inspectable control flow - the right tool when an agent must run for minutes to hours, survive restarts, and gate on human input. For TypeScript-first frontend builders the Vercel AI SDK is lighter weight; LangGraph wins when durability and complex branching matter more than UI integration. Compute-heavy LangGraph nodes can be offloaded to platforms like Modal Labs or sandboxed via E2B Code Sandboxes.
Timeline
- 2026-07-04 | Linked LangSmith Sandboxes and the Auth Proxy into the LangChain product map: LangGraph owns orchestration, while LangSmith supplies trace/eval/deploy plus sandbox credential and egress controls. Source: LangSmith docs, 2026-07-04
- 2026-06-18 | Updated with official docs positioning: LangGraph as orchestration runtime, Deep Agents as harness, LangSmith as trace/eval/deploy layer; linked into the new agent runtime architecture. Source: LangGraph docs, 2026-06-18
- 2025 | LangGraph reached its 1.0 milestone, hardening the durable-execution and checkpointing model and decoupling further from LangChain. Source: langchain-ai/langgraph, https://github.com/langchain-ai/langgraph, 2026-05-31
- 2026-05-31 | Tool page created. Repo at 33,465 stars, active daily, described as a "low-level orchestration framework for building stateful agents," trusted by Klarna, Replit, and Elastic. Source: langchain-ai/langgraph README, https://github.com/langchain-ai/langgraph, 2026-05-31