Checkpointing

Save workflow state

A checkpoint is a durable record of where a run got to and what state it had accumulated. After a crash, a restart or a deliberate pause, the workflow resumes from there instead of starting over.

This is not memory and not conversation history. It is narrower and more boring: which step finished, and what it left behind.

Any workflow that pauses for a human needs one, because otherwise the pause and a crash are the same event. It also changes the economics of long agent runs — without checkpoints a failure at step nine means paying for steps one through eight a second time.