State Persistence
The ability to save where an agent run is and resume it later, surviving crashes and spanning separate requests.
State persistence is the framework saving the full state of an agent run, so it can be paused and picked up again. Without it, an agent lives only for the duration of one request; if the process restarts or the run takes longer than a single call, the work is lost.
It enables three things that matter in production. First, durability: a long agent run can survive a crash or deploy and resume where it left off. Second, human-in-the-loop: a run can stop for approval, persist, and continue later. Third, rollback: with saved checkpoints, you can return to an earlier step and try a different path.
LangGraph built its checkpointer around exactly this and treats persistence as core. Whether you need it depends on your agent: a quick request-response agent that finishes in seconds may never need persistence, while a long-running or resumable workflow depends on it. If you anticipate either, prefer a framework that makes persistence easy rather than bolting it on later.