Runtime
A graph that can wait for a person.
LangGraph models agents as steps in a graph that share one saved state. That makes pausing for a human a normal step, not a special case.
Each agent is a node. Arrows between them are edges, and a conditional edge picks the next step from the current state, such as "Wi-Fi problem, go to the fixer".
A checkpointer saves the state after every step. When a node calls interrupt, the run stops and waits. A person can answer hours later, and the run resumes from that exact point.
Send starts many copies of a branch at once, which is how one batch becomes parallel work.
- Node
- One agent or step
- Conditional edge
- Routing on the state
- Send
- Parallel fan-out
- interrupt
- Pause for a person
- Checkpointer
- Saved state and history