A good fit for
- → Long-running agents
- → Approval-and-resume workflows
- → Stateful orchestration with testable transitions
Know the limits
- → Checkpointing alone does not make external writes exactly-once.
- → Durability modes trade throughput for crash recovery guarantees.
- → Graph abstractions add complexity to tasks that need only a simple loop.
Your first useful workflow
Define state
Separate user intent, observations, pending actions and completed effects.
Choose durability
Match checkpoint mode to the cost of losing an intermediate step.
Kill and resume
Test a crash immediately before and after an external write.
The practical take
Persistence lets a workflow resume. Idempotency prevents its resumed steps from charging twice.
Sources & evidence
Capabilities above are summarized from these official documents. Workflow suggestions are editorial guidance, not a claim of measured product superiority.