Checkpointing (Agent State)

Persisting a snapshot of an agent's full execution state — conversation, intermediate results, position in the workflow — at each step, so the run can be paused and later resumed, inspected, or rewound without loss. Checkpointing is the mechanism beneath Durable Execution: it's how an agent survives a crash or waits days for a human at an Approval Gate and then continues from exactly where it stopped. It also enables 'time-travel' debugging — replaying a run from an earlier checkpoint to see what a different decision would have done — and clean Human-in-the-Loop (HITL) interrupts. LangGraph exposes checkpointing as a first-class primitive (its `interrupt()` + checkpointer); other frameworks provide equivalents or lean on external durable-execution engines. If your agent takes irreversible actions or waits on async human input, checkpointing is what keeps 'pause and resume' from meaning 'restart and re-run.'