Checkpoint a workflow and resume after failure

Restarting a workflow from the beginning can repeat tool calls or external writes. A checkpoint records what completed, what failed and which step may safely resume.

Resume without repeating work

The agent controls lab saves an authored failure after validation and lookup.

from agent_cases import checkpoint_case

result = checkpoint_case()
assert result["resume_from"] == "draft"
assert result["repeated_completed_steps"] is False
assert result["resumed_steps"] == ["validate", "lookup", "draft"]

The checkpoint includes the lookup idempotency key and temporary timeout. Recovery advances to draft instead of rerunning completed steps.

Persist recovery-critical state

Store run/config version, validated inputs, completed steps, tool receipts, idempotency keys, approvals, next step, attempt budgets and last error. Write checkpoints atomically after a step’s durable effects are known.

Not every step is replay-safe. Mark compensation or verification actions for uncertain outcomes. If a network timeout occurs after a write request, query by idempotency key before retrying.

Test every interruption boundary

Fail before and after each side effect, serialize state and resume with a fresh process. Confirm output equivalence, one side effect and preserved total budgets. Reject resume when code/state versions are incompatible without a migration.

The Generative & Agentic AI course connects checkpoints to idempotency, cancellation and partial-completion runbooks.

Exercise

Checkpoint a four-step workflow after every transition. Inject failures at eight boundaries and prove the recovered result matches a clean run without duplicated writes.

Continue learning

This article is part of the Agent workflows and state sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Design typed state for a multi-step workflow.
• Continue with Idempotency keys for agent tool calls.

Reference: LangGraph persistence.