Skip to content

Durable runs

Checkpoints, interrupts, streaming, and the boundaries that keep agent work recoverable.

Updated View as Markdown

Durability is explicit: a graph persists caller-owned state through a CheckpointStore, while an application selects durable stores for pending deep-agent runs, todos, and subagent tasks. In-memory stores are reference implementations for tests and examples.

An Interrupt pauses a graph at a known point and returns a ResumeToken. Resume supplies a caller-owned StateUpdate, preserving the thread identity and checkpoint history. Streaming exposes state, updates, retries, checkpoints, interrupts, debug events, and custom node events through Flow.Publisher.

Use durable storage in production whenever a run must survive process restarts or wait for a human decision.

Navigation

Type to search…

↑↓ navigate↵ selectEsc close