What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
When a background AI process works in a demo but fails as soon as it has to wait, the problem is often not the model call itself. The demo may keep one request, process, and in-memory state alive; production adds timeouts, worker restarts, retries, delayed external services, and human approval. The fix is to separate waiting for a response from preserving and resuming the whole workflow.
Why a background AI job can fail after the demo
A demo often makes progress appear continuous: one process starts the work, keeps its state in memory, waits for the model or a tool, and returns the result. A production system may stop waiting when a client connection times out, kill a worker during a deploy, or retry a job after a temporary error. A human approval step can stretch the workflow from seconds to hours or days.
As an Amazon Associate I earn from qualifying purchases.
Those are separate failure modes. An asynchronous model request can free a client from holding one connection open, but it does not automatically preserve the surrounding workflow: what steps already finished, which external actions happened, whether approval is pending, or where to continue after a restart.
First decide what needs to survive the wait
Before choosing an API or orchestration tool, trace the job from start to finish and identify what must still be true when it resumes. Write down the model calls, tool calls, database changes, external actions, approval gates, and final result. For each point where the work can pause, ask what would happen if the process disappeared at that moment.
#1 Best Overall
- Only one long model response: the key need may be to submit the request asynchronously and check its status later.
- A conversation that continues across turns: the application needs a way to recover the relevant conversation state after a request or worker ends.
- A multi-step job with external effects, long waits, or approval: the application needs durable progress and a defined way to resume or retry each step.
This distinction prevents a common mistake: adding polling to a long model call and assuming that makes a multi-step agent workflow durable.
Choose where the waiting and state live
| Approach | What it addresses | What your application still needs | Best fit |
|---|---|---|---|
| Provider background request | Runs a supported model response asynchronously and lets the application poll its status. | Handling the terminal result, persisting any broader workflow state, and recovering surrounding steps after a restart. | A single long model response where the application can manage the remaining lifecycle. |
| Application-managed or persisted agent state | Preserves conversation or continuation state outside a short-lived process. | Resuming the right step, coordinating workers, and making external actions safe to repeat. | Conversational agents, resumable turns, or approval flows where the application needs control of state and storage. |
| Durable workflow orchestration | Tracks multi-step execution across waits, retries, and process restarts. | Defining correct step boundaries, error handling, and safe behavior for external side effects. | Jobs with long waits, human approval, worker restarts, or several retryable steps. |
For one long model response, use background execution
OpenAI documents background mode for long-running work with the Responses API: start a response asynchronously, poll while it is queued or running, and handle its terminal status and output. That solves the problem of keeping a client request open for the duration of the model call; it is not by itself a durable record of every step in an agent workflow. See the OpenAI background mode guide.
For resumable turns, make continuation state explicit
State can be held as application-managed history, in a persisted SDK session, or through a server-managed conversation identifier. Choose based on who controls storage, whether state must be shared across workers, and whether the application must resume an approval or turn later. An in-memory session is useful during a demo but disappears with its process; it should not be treated as persistent state. The OpenAI Agents runtime guide describes these continuation approaches.
For multi-step work, use durable orchestration when recovery warrants it
If a job must survive long waits, retries, or restarts, a workflow engine can persist progress and coordinate resumption rather than relying on one worker to stay alive. The OpenAI Agents SDK documents integrations for Dapr, Temporal, Restate, and DBOS in its durable agent integrations guide. Temporal’s OpenAI Agents integration runs model calls as Activities so workflow replay does not repeat those calls. These capabilities address execution recovery; they do not guarantee that application logic or external business actions are correct.
Design each step to recover safely
Treat each model call, tool call, database write, and external API action as a step with a clear boundary. For every step, decide what is persisted before it begins, what counts as success, and what the workflow should do if the worker stops before recording the outcome.
- Set a timeout: choose how long the step may run before the system classifies it as timed out. Keep this separate from the overall job deadline and any client-facing wait.
- Classify failures: distinguish transient infrastructure problems from errors that need a correction, user input, or a policy decision. Do not automatically retry every failure.
- Bound retries: set a maximum attempt count or deadline and a backoff policy. OpenAI’s deployment checklist advises checking status and error codes, honoring
Retry-Afterwhen supplied, and otherwise using bounded exponential delays with jitter for appropriate transient errors. - Persist progress at meaningful checkpoints: record enough information to continue without repeating completed work. Temporal documents activity heartbeats as a way to carry checkpoint data across retries in its task behavior guide.
- Make external effects repeat-safe: use idempotency keys or deduplication where supported, or reconcile the external system’s state before repeating an action. A retried step may run again even when the orchestration layer durably schedules it.
- Route non-retryable outcomes deliberately: surface validation failures, policy blocks, and unresolved human decisions for correction or review instead of redispatching them as if they were transient.
Make approval a persisted workflow state
A human approval wait is not just a slow API call. Save the proposed action, the information the reviewer needs, the workflow’s current step, and a durable identifier that lets the approval response resume the correct job. When approval arrives, validate that it still applies to the saved action before continuing. If the process restarts while waiting, it should recover the pending approval rather than create a second request or lose the job.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Check the system before calling it production-ready
Test the failure cases the demo concealed. Stop a worker during a model call, after a tool returns, and after an external action succeeds but before the workflow records success. Also test a slow provider response, a temporary network error, an exhausted retry limit, a delayed approval, and an approval that arrives after the job has expired.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problems- Can another worker find the job and its continuation state after a restart?
- Can an operator see the current step, last error, attempt count, and next scheduled action?
- Can a retry avoid duplicating a payment, message, record, or other external effect?
- Does the workflow stop or request intervention when a failure is not safe to retry?
- Are job deadlines and data-storage choices appropriate for the workflow?
Choose the lightest design that passes those checks. Compare expected wait duration, recovery requirements, state ownership, operational burden, observability, fit with your existing stack, and cost. The documentation describes capabilities, not a universal best tool; durable execution can preserve and resume work, but the application remains responsible for valid outcomes and safe side effects.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
