Durable Execution, Checkpoints & Recovery
Can the workflow resume after interruption without forgetting decisions or repeating external actions?
Reference notes by Min · Reviewed October 15, 2026
Definition
Durable agent execution preserves sufficient workflow state to continue after process failure or interruption. Checkpoints record progress, while recovery determines which work can safely resume and which outcomes must first be reconciled.
Long-running agents may cross deployment changes, human review delays and dependency outages. A useful checkpoint records facts that remain meaningful after the process disappears: task identity, accepted results, pending operations, authorization references and the next permitted transition.
What it covers
- State
- Record completed steps, pending actions, evidence and relevant versions.
- Boundaries
- Persist state around consequential operations with explicit transaction semantics.
- Replay
- Distinguish reconstructing workflow state from repeating a model or tool call.
- Validity
- Recheck expired approvals, permissions and changed external conditions.
Persisting an intention and committing an external action are different events. A crash between them creates uncertainty unless the surrounding protocol can reconcile the result. Durable execution needs a clear treatment of that gap, not merely a retry policy.
In practice
Practical design patterns by Min.
- Keep authoritative task state separate from a lossy conversation summary.
- Use stable operation identifiers and downstream idempotency where available.
- Represent uncertain outcomes explicitly and reconcile before retrying writes.
Version the workflow and its relevant artifacts. A resumed task may have been created under an older prompt, tool schema or policy. Decide whether it can continue under the original version, needs migration or requires renewed review. Silent changes make incident reconstruction difficult.
Research notes
Source-grounded details and editorial applications. Follow each link for the original document.
Workflow persistence is a runtime capability
Temporal documents workflow execution and durable progress. Agent designers still need explicit handling of external effects, approvals and nondeterministic model results within their chosen runtime.
Coordinating state and messages has limits
AWS describes the transactional outbox pattern for coordinating database changes with message publication. Downstream consumers still need appropriate duplicate handling; the pattern is not a universal exactly-once transaction across services.
Source: AWS Prescriptive Guidance: Transactional outbox pattern ↗
Working example
An agent crashes after submitting an order
The supplier may have accepted the order, but the local workflow has no recorded confirmation.
- Resume in an unknown-outcome state rather than treating the order as failed.
- Query the supplier using the original operation or order identifier.
- Record the confirmed outcome or escalate if it cannot be established; do not blindly create a second order.
Why the controls matter
The order example demonstrates why recovery needs an external reference. A model can infer that an order probably succeeded, but that inference does not establish the supplier’s state. Use service evidence and provide a human path when it remains unavailable.
Common pitfalls
- Saving chat history and assuming it fully describes external state.
- Replaying a non-idempotent operation because the previous process died.
- Resuming with an approval whose scope or validity has changed.
Retention and cleanup also matter. Completed checkpoints can contain sensitive data or references to revoked credentials. Define access, deletion and archival rules, and prevent retired workflows from restarting automatically through an old queue or scheduled retry.
Questions for operational review
Use these questions to decide what evidence your team needs. They are editorial prompts, not a compliance checklist.
- Which state survives a worker failure?
- How is an ambiguous external effect reconciled?
- What is revalidated before an interrupted task resumes?
Articles linking here
References
Primary documents reviewed for this entry. Examples and implementation patterns are editorial synthesis.
- Temporal: Workflow Execution overview ↗Accessed October 15, 2026
- AWS Prescriptive Guidance: Transactional outbox pattern ↗Accessed October 15, 2026