Skip to content

Before Your AI Tries Again

Architecture2026-11-026 min read

A timeout after launch does not establish that an operation never ran; execution and result recovery require separate evidence.

When an AI-assisted workflow stops answering, your operations team still has a decision to make. Did the work stop, or did its confirmation disappear? If the next step updates a supplier record or sends a customer message, starting again can repeat an action the team is trying to finish. Those are illustrative workflows, and the operations lead who owns delivery must decide whether another attempt is allowed.

Before authorizing another attempt, ask what the original operation did and who is responsible for resolving an uncertain outcome. The person waiting for delivery needs that answer as much as the engineer operating the system. A timeout gives you a reason to investigate. By itself, it does not establish that the first attempt never ran.

I ran into this in my own operation while coordinating an AI review. Our caller stopped waiting, but the original review later finished. We recovered its result without asking the model to do the work again. The useful lesson was the separation between running an operation and finding out what happened to it.

Give the uncertainty an owner

Choose one AI-assisted workflow where a repeat would matter. Name the person waiting for its result and the person allowed to decide what happens after a missing confirmation. They might be the same person. Make that responsibility explicit before the first ambiguous failure arrives.

For the supplier-record illustration, an operations lead might be waiting for a completed update while an engineer sees a request timeout. The engineer needs to establish whether the update reached the receiving system. The operations lead needs to know whether another attempt would repeat a change or simply retrieve its status. Neither question is answered by the caller's error alone.

Ask what the workflow can distinguish: an operation that never started, one still in progress, one with a retained result, and one whose outcome remains unknown. These are questions to investigate in the actual system, not labels to assign from a message on the screen. A missing result can remain unknown even after the worker has stopped.

Your team also needs a rule for that last condition. Decide who investigates, which evidence they can read, and when they should stop and escalate. Leaving uncertainty visible can be more useful than turning it into a confident failure and giving the next worker permission to start over.

What happened after our timeout

During an internal qualification on October 23, our coordination work sent a review to an AI worker. The caller waited for a receipt and timed out after 180 seconds. That established how long our caller waited. It did not establish that the native review had been cancelled or had never started.

The original operation later completed. We used a separate read-only lookup, bound to that operation and the digest, a fingerprint of its input text, to collect the retained result. The lookup did not initiate another inference. The original review's record still showed one dispatch attempt.

The surrounding stored records mattered. We retained the original intent and its completion context, and the worker's stored reply digest matched the digest of the reply delivered to our coordinator. The coordinator then checked the requested deliverable before closing the task. Recovering a reply and accepting the work were separate steps.

This was a bounded internal review task. It was not a customer update, and it did not measure duplicate charges or customer harm. It showed that a lost or late receipt could be recovered for this operation without repeating its inference. It did not establish a general delivery guarantee for every agent or failure.

We also kept the failed wait in the record. Replacing it with a clean success would erase the distinction that made the recovery worth inspecting: our first observation was incomplete, and the later evidence answered a question the timeout had left open.

Recover the original operation

An engineer needs an identity that survives the lost answer. Keep the original operation identifier with the intended input and the receiving system's reference, where one exists. If every recovery attempt invents a new operation, your team can lose the connection between the work that might already have happened and the work it is about to authorize.

In our implementation, result inspection is a separate path from execution. The inspection request identifies the original operation, supplies the input digest and explicitly requests read-only handling. The reader checks the stored identity and digest before returning the retained result. A conflict is refused; missing or unusable state remains unknown.

That separation matters because a function called "check status" can still have side effects. Follow its consuming path. Does it only read retained evidence, or can it enter the worker again when a file is missing? In our path, inspection does not call the model. The reservation for an existing operation is preserved rather than deleted to make another attempt possible.

A digest answers a limited question about bytes. It helps bind a retained reply to the intended request; it does not establish that the reply is correct or that the requester has authority. The storage boundary and the identities your system trusts still matter. Our evidence assumes cooperating processes under the same operator, not protection from a hostile process with that operator's access.

The same limit applies to an idempotency key. Where a receiving service supports one, inspect its actual scope, retention and behavior for changed input. Reusing an identifier is useful only when the relevant execution path enforces the intended rule. An identifier written in a document cannot prevent duplicate effects at a service that ignores it.

Some workflows already have an authorized retry contract that handles these cases. Use its documented boundary. The concern is an automatic repeat justified only by a timeout, with no account of the original attempt. More transport attempts and another business action are not necessarily the same thing.

A result still needs acceptance

Finding a matching result closes one question. Your delivery owner still needs to know whether it satisfies the requested work. An AI can return a review that addresses the wrong material, misses a required check or stops short of the deliverable. Identity and content quality need different evidence.

If the workflow also changes an external system, inspect that effect separately. A retained model answer does not prove that a supplier update was applied or that a customer received a message. Conversely, the external effect may be present even when the caller has no final answer. Keep the intended result, observed effect and acceptance decision distinct.

In our review case, the coordinator checked the deliverable and correlated the returned evidence before recording completion. That responsibility did not move to the model because a reply arrived. The person accepting the work still owned the decision to close it.

Your handoff should tell the next operator what is known, what remains uncertain and which action is permitted. Preserve the original failure and later readback together. A historical success is useful context, but the next operator must still check the current operation rather than borrow certainty from an earlier run.

Make the next attempt a decision

Ask your engineer to walk through one safely retained case, or an already authorized harmless test, where the caller loses its confirmation after work starts. Trace the original operation to the receiving side and inspect what recovery actually does. This is an inspection aid, not permission to interrupt a live customer workflow.

Then ask what happens when the evidence is unavailable. A recovery path should have a bounded wait and a named owner for escalation. If the first attempt's outcome cannot be resolved, keep that uncertainty in the decision. A human approval for another attempt does not prove the first one had no effect.

The owner may decide to reconcile the receiving system, wait for a late result, stop the workflow or authorize a further attempt under an understood duplicate risk. Engineering should make those choices operable, including any service-specific retry protections. The important change is that the next action follows the known state and authority, rather than the emotional pressure to make an error disappear.

I want an AI workflow's handoff to survive a missing answer. Someone should be able to identify the original attempt, inspect the evidence it left and explain who can authorize the next move. Recovering that account can be the work that finishes the workflow.

Which AI-assisted workflow in your operation needs the first attempt accounted for before it can run again? Describe that workflow in the comments.

Share this noteLinkedInX