When to reach for it
- Agent runs are lengthy or expensive.
- External systems might temporarily fail.
- Manual review between steps is required.
Checkpointing.
Save state. Resume cleanly.
The execution state is persisted at well-defined boundaries (per node or super-step) so an interrupted or failed run resumes from the last checkpoint instead of restarting — the durable-execution substrate that also enables human-in-the-loop pauses.
In practiceA multi-hour data-migration agent checkpoints after each table, so a network blip resumes from the last table instead of restarting.
When it backfires
- Steps are atomic and cheaply repeatable.
- The state unnecessarily persists sensitive data.
- Resuming makes no business sense.
The tradeoff
Higher fault tolerance is gained against storage, privacy, and consistency overhead.
What it actually does.
State is saved so a crashed run resumes mid-flight.
Two ways this pattern will hurt you.
Checkpoints hold secrets
The checkpoint serializes the full state including API keys and user PII. A resumed run leaks sensitive data to logs.
Fix · Redact secrets before serialization. Store sensitive fields in a vault and reference them by handle in the checkpoint.
Resumption with stale tool state
The agent resumes from a checkpoint but the external world has changed. The tool state in the checkpoint is now invalid.
Fix · Invalidate tool-related state on resume. Re-run discovery or validation before continuing tool-dependent branches.
Where Checkpointing is native.
Microsoft Agent Frameworkdurable run stateNative
Next pattern