[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-agent-checkpointing::en":3,"gloss-cluster-agent-checkpointing::en":26,"gloss-next-agent-checkpointing::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"agent-checkpointing","agents","Agent Checkpointing","Agent checkpointing is persisting an agent's state at defined points in a run so the run can be resumed rather than restarted. The state worth saving is more than a conversation transcript: the step counter and the budget consumed so far, the tool calls already made and their results, whatever the agent has committed externally, and the pending plan. A run that cannot be resumed has to be re-executed from the beginning after any interruption, and re-executing an agent is not free. It costs the tokens again, and it repeats every side effect the first attempt already caused. That second cost is why checkpointing and idempotency are one design problem seen from two angles. Resuming safely requires knowing which external actions have already happened, and the reliable way to know is to record the intent to act before acting, mark it complete afterwards, and give each action an idempotency key so a repeat is absorbed rather than duplicated. Without that record, a resumed agent sends the second invoice. Checkpoints also change what an operator can do. A run paused at a checkpoint can be inspected, edited and continued, which is how human approval fits into a long task without a person watching it end to end, and a failed run can be re-driven from the last good point with a corrected input instead of being thrown away. The mechanism costs storage and write latency, so checkpoint at semantically meaningful boundaries rather than after every model call: once a tool result is stored, once a plan step completes, and before anything irreversible. Replay cannot reconstruct the state exactly, because model output is not deterministic, which is precisely the reason the state is stored rather than recomputed.","Agent checkpointing saves run state so an interrupted agent resumes instead of repeating its side effects, and why that makes idempotency part of the design.",null,[11,14,17,20,23],{"slug":12,"name":13},"agent-budget","Agent Budget",{"slug":15,"name":16},"agent-loop","Agent Loop",{"slug":18,"name":19},"agent-memory","Agent Memory",{"slug":21,"name":22},"agent-trajectory","Agent Trajectory",{"slug":24,"name":25},"idempotency","Idempotency",[27,29,32,33,34,38,39,42,45,48,51,55],{"slug":12,"category":5,"name":13,"updated_at":28},"2026-08-24T02:46:37+00:00",{"slug":30,"category":5,"name":31,"updated_at":28},"agent-handoff","Agent Handoff",{"slug":15,"category":5,"name":16,"updated_at":28},{"slug":18,"category":5,"name":19,"updated_at":28},{"slug":35,"category":5,"name":36,"updated_at":37},"agent-sandbox","Agent Sandbox","2026-08-24T02:46:38+00:00",{"slug":21,"category":5,"name":22,"updated_at":37},{"slug":40,"category":5,"name":41,"updated_at":28},"agentic-rag","Agentic RAG",{"slug":43,"category":5,"name":44,"updated_at":37},"computer-use","Computer Use",{"slug":46,"category":5,"name":47,"updated_at":28},"context-engineering","Context Engineering",{"slug":49,"category":5,"name":50,"updated_at":37},"dry-run","Dry Run",{"slug":52,"category":5,"name":53,"updated_at":54},"escalation-path","Escalation Path","2026-08-24T03:30:02+00:00",{"slug":56,"category":5,"name":57,"updated_at":28},"human-on-the-loop","Human-on-the-Loop"]