Agent Checkpointing
A long agent run is a fragile thing to hold only in memory: a process restart, a routine deploy, or a reviewer who takes an hour to approve something will each destroy it. Agent checkpointing is the discipline of writing the run's state to durable storage after every step, so it can be picked up exactly where it stopped.
What a checkpoint holds
Enough state for the next decision to be made as if nothing had happened:
- The message history, plus any plan or scratchpad the agent is working from.
- Results already returned by tools, so completed work is not repeated.
- The position in the loop — which step comes next, and what is waiting on approval.
In LangGraph this is a checkpointer backed by SQLite or Postgres, writing a record per node under a thread id; durable-workflow engines such as Temporal apply the same idea to an entire run.
What it makes possible
Checkpointing turns a chat-shaped loop into something operable. A crashed run resumes instead of restarting, so the tokens and tool calls already paid for are not spent twice. A run can sit blocked on an approval gate for hours without holding a process open. And because states are stored in sequence, an engineer can replay a trajectory to find where a plan drifted, or rewind to an earlier checkpoint and re-run from there with a corrected instruction.
Why it matters
Without checkpoints the practical ceiling on an agent is one uninterrupted session, and any failure costs the entire task. The price is discipline: state must be serialisable, and a checkpoint taken after an action already executed must not re-execute it on resume.
Agent checkpointing is the practice of persisting an agent's state — conversation history, plan, intermediate results, position in the loop — at defined points during a run. The saved snapshot lets execution resume from the last good step instead of restarting after a crash, timeout, or deployment. For long-running agents, checkpointing is what makes progress durable: without it, hours of tool calls and tokens are lost to a single interruption.
Think of saving your progress in a long video game. If the power cuts out you do not start the whole game again — you reload the last save point. Agent checkpointing is that, for software working through a long task on its own.
These tasks can run for hours and involve hundreds of small steps. Computers restart, connections drop, and sometimes the software has to wait for a person to approve something. So after each step it quietly saves a record of what it has done and what comes next.
When it starts up again it carries on from that save point, instead of repeating work that was already finished — and already paid for.