Agent runtime

LangGraph

LangChain's library for building stateful agent workflows as graphs — nodes, edges, checkpoints, and human-in-the-loop interrupts — one of the common substrates for production agent orchestration.

Operating principle

Production AI is a system: context, tools, permissions, traces, evals, and feedback loops around the model.

What it is

LangGraph models an agent workflow as an explicit graph: nodes do the work, edges route between them, and a checkpointer persists state so a workflow can pause, resume, and survive restarts. It supports interrupts for human approval, sub-graphs for delegation, and streaming of intermediate state.

Why it matters

An agent loop that lives in a while-loop doesn't survive production: no restartability, no visibility, no place to attach an approval gate. Making the control flow an explicit, checkpointed graph is one credible answer — the supervisor and sub-agent patterns most teams reach for are documented LangGraph patterns.

How we use it

LangGraph is one of the substrates we extend when a client already runs it, alongside the OpenAI Agents SDK and durable-execution engines like Temporal. Our own default is Pydantic AI with a deliberately simple stage-and-gate executor; we add the same runtime contract — versioning, evals, traces, gates — on either.

Related resources