AuxeonChoose an AI specialist
Menu
← All insights

Long-running AI agents need a recovery plan

An agent may work for hours and still leave the business unsure what happened. Design how it resumes, checks prior actions, and proves completion before extending its run time.

Share on LinkedIn ↗

OpenAI’s October 14, 2026 case study of Third Moment Research describes agents conducting investigations lasting 36–48 hours. The detail worth noticing is the surrounding system: when a job stops unfinished, another supervisory layer examines the run and decides whether to retry, stop, or seek human review.1

This is a vendor-published account of one research environment, not a reliability benchmark for your operation. It raises a practical operating question: if an agent loses its connection halfway through your workflow, what happens next?

What should happen when an AI agent is interrupted?

The workflow should preserve what is known, identify what remains uncertain, and resume only from a verified state.

Consider a hypothetical service-request process. An agent creates a customer case, attaches supporting documents, and prepares a confirmation for approval. The connection fails just after the case-creation request. Starting again could create a duplicate case. Declaring success could leave the documents unattached. Neither follows from the timeout alone.

The business needs a third state: outcome unknown. That state should trigger a check of the receiving system or a named person’s review, rather than an automatic guess.

Where should an agent’s progress be recorded?

Keep a durable task record outside the agent’s current conversation. Record the requested outcome, completed steps, pending decisions, and references to results in the systems where the work happened. Retain only the information the workflow needs, under the business’s access and retention rules.

Anthropic’s November 2025 engineering report offers relevant background. Its coding-agent experiments used progress records, version history, and explicit feature tests to help later sessions continue the work. The report also describes agents declaring completion too early. Those findings concern a particular coding setup, not every business agent.2

For a service workflow, the equivalent record might contain the stable operation identifier, the case identifier if known, the intended effect, the last attempted action and time, attachment status, approval reference and current validity, cancellation status, evidence links, and the next unresolved step. The next worker should be able to recover the task from that record and the receiving systems, without reconstructing its history from a chat transcript.

When is it safe for an agent to retry?

A retry is appropriate when the action is still authorized and the system can control its effects.

AWS’s engineering guidance explains how an idempotent API can recognize a repeated request using a caller-supplied identifier, avoiding an additional side effect. It also describes the uncertainty created when a request receives no response.3

Ask the implementer which connected actions actually support that protection, where it is enforced, how long duplicate identifiers are retained, and whether every retry reuses the same identifier for the same intended operation. A task identifier in an agent’s notes does not prevent a duplicate unless the connector passes it through and the receiving application enforces it. If that protection is unavailable or has expired, reconcile the unresolved outcome before another attempt.

The execution platform matters too. AWS distinguishes Standard, Asynchronous Express, and Synchronous Express Step Functions: its documentation describes exactly-once workflow execution for Standard, at-least-once execution for Asynchronous Express, and at-most-once execution for Synchronous Express, with further qualifications around retries and retained state.4 Those are workflow-execution guarantees, not a promise that every downstream API call, message, or external side effect occurs exactly once. Do not accept a blanket “exactly once” claim without tracing the complete workflow and each connector.

What proves the task is complete?

Define completion in terms the receiving business can verify.

For the hypothetical service request, completion might mean one case exists, the required documents are present, and the approved confirmation is recorded as sent. If the service only confirms that a message was accepted for sending, report that narrower status. Do not label it delivered or read.

Use separate states for prepared, awaiting approval, attempted, confirmed, and unresolved. Link the evidence needed for each transition. A confident final paragraph from the model is not the acceptance test.

A restart neither expands the agent’s permissions nor cancels an authorization that remains valid. Before the next consequential action, the system should re-check that the current permission and approval still cover that specific action, and that neither has expired, been revoked, or been superseded by a cancellation. Cancellation must be durable and checked at the action boundary. If an action may already be in flight, record the outcome as unresolved and reconcile it rather than assuming cancellation reversed it. Existing permissions, budget limits, and the agent’s mandate must still apply.

How should a business test recovery before launch?

Use an isolated test environment and approved sample data. Ask the team to demonstrate:

  1. An interruption immediately before and immediately after a simulated write.
  2. A delayed response where the receiving system has already completed the action.
  3. A restart while human approval is pending.
  4. A cancellation recorded before the next external action, including a race in which a prior request may still be in flight.
  5. A task that remains incomplete when its agreed time or cost limit is reached.
  6. A retry through the real connector using the same operation identifier, followed by the same identifier with changed intent. Verify one side effect for the duplicate request and no silent acceptance of the changed request.

For each case, inspect the receiving system as well as the agent’s account of events. Check for duplicate records, missing work, bypassed approvals, and misleading completion messages. Record who owns unresolved outcomes and what evidence lets that person resume or close the task.

These are proposed acceptance tests, not reported Auxeon customer results. Set the pass criteria for your workflow before testing, and repeat the relevant cases after material changes.

Longer run time gives an agent more room to work. A recovery plan gives the business a way to verify what happened and resolve what remains when the run ends.

Sources

  1. OpenAI, October 14, 2026. How Third Moment Research runs 10,000 agents for 36+ hours. Vendor-published customer case study.↩
  2. Anthropic, November 26, 2025. Effective harnesses for long-running agents. Background engineering research, outside the current news period.↩
  3. Amazon Builders’ Library, Malcolm Featonby. Making retries safe with idempotent APIs. Undated engineering guidance, accessed October 20, 2026.↩
  4. AWS Step Functions Developer Guide. Choosing workflow type in Step Functions. Living product documentation, accessed October 20, 2026.↩
Related pageAuxeon Systems Review