Use retry logic for a bounded, classified transient failure when repeating the operation is safe. Use durable recovery when a multi-step agent workflow must survive a crash, restart, out-of-memory termination, or redeployment. Most production agents need both: retry eligible local operations, checkpoint progress between stages, reconcile uncertain side effects before repeating them, and send persistent failures to a fallback or a person.
A retry repeats an operation that a running process knows has failed. Recovery resumes a durable attempt after the process handling it disappears. Confusing those cases can create duplicate messages, lost work, retry storms, or workflows that restart from the beginning. The practical rules below follow guidance from OpenAI’s Agents API error guidance and the AWS Agentic AI Lens.
As an Amazon Associate I earn from qualifying purchases.
The decision in one view
| Situation | Preferred mechanism | Guard you need |
|---|---|---|
| A timeout, rate limit, or temporary overload on a request that has not completed | Bounded retry | Classify the error, use exponential backoff with jitter, honor Retry-After, and stop at an attempt or time limit. |
| Invalid arguments, expired credentials, a rejected permission, or a billing/quota constraint | Do not automatically retry | Correct the input or account condition, then submit a new operation deliberately. |
| A workflow spans several stages and must continue after a process disappears | Durable recovery | Persist checkpoints, stage results, or watermarks that a newly started handler can read. |
| A tool may have completed an external action before the client saw an error | Reconciliation before retry or recovery | Query the provider or your operation record; use an idempotency key or equivalent deduplication rule. |
| A dependency remains unavailable after the retry budget is exhausted | Fallback, degraded mode, or human review | Record the failed stage and context in traces so the next action is informed. |
The dividing line is not whether an application is called an “agent.” It is whether the failed work is a safe, repeatable operation or a durable workflow with progress worth preserving.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What retry logic does
A retry is a bounded repetition
Retry logic belongs to an operation that is still being managed by a running handler. The handler receives an error, classifies it, and decides whether another attempt is reasonable. A timeout caused by temporary congestion may qualify. A malformed tool argument does not. A retry policy should always have a finite attempt count, a deadline, or both; otherwise an agent can consume resources indefinitely while the underlying condition remains unchanged.
#1 Best Overall
- [Ultra-Lightweight Ear Buds Designed for Small Ears] Each earbud weighs only 3.7g and features a compact, ergonomic in-ear design made especially for small ears. Secure, low-profile, and comfortable for workouts, all-day wear or overnight sleeping.The ultra-slim profile sits flush in your ear with no hard edges poking or rubbing, so you can even sleep on your side without discomfort.
- [Immersive Stereo Sound with TOZO OrigX Technology] TOZO OrigX tuning delivers clear vocals, balanced mids, and natural stereo sound for music, podcasts, and videos.
- [Long Battery Life for Daily Use] Get up to 7 hours of playtime on a single charge, with up to 32 hours total using the charging case—ideal for workdays, commuting, and extended listening sessions.
- [Bluetooth 5.3 & Stable Connection] Bluetooth 5.3 provides fast pairing, stable wireless performance, and reduced dropouts as you move around home or office.
- [Deep Bass with Clear Vocals] High-performance drivers produce punchy bass while keeping vocals clean and detailed for everyday listening.
Before repeating a request, determine whether it actually ran and what actions completed. An HTTP connection can fail after a service accepted the request, so “the client saw an error” is not proof that nothing happened. OpenAI’s Agents API guidance separates errors that occur while submitting a request from failures after work has started and advises checking the operation’s outcome and completed actions before retrying. It also recommends bounded behavior, honoring a server’s Retry-After instruction when present, and stopping automatic retries when the error changes or the limit is reached: Agents API errors and recovery.
Classify before you retry
- Transient and likely to clear: temporary overload, a rate limit, a network timeout, or a short-lived service-unavailable response. These can be candidates for retry when replay is safe.
- Permanent until something changes: invalid JSON or schema, a missing required field, an expired credential, a denied permission, or a billing and quota restriction. Repeating the same request will not repair the cause.
- Unknown outcome: the connection failed after a tool was invoked or a stream began. Check the provider, a durable operation record, or a queryable result before deciding whether to repeat.
- Process loss rather than a reported operation error: the worker crashed, was killed, or was redeployed. This is a recovery problem if the workflow has durable state.
Backoff is part of correctness
Immediate, identical retries can amplify an outage. Use increasing delays and random jitter so many agents do not resubmit at the same instant. A service-provided Retry-After value takes precedence when applicable. Carry the retry budget across nested calls: if an agent retries a model call and the model call retries a tool call, each layer should not independently turn a small budget into an unbounded multiplication of attempts.
When the limit or deadline is reached, return a classified failure to the workflow. Do not silently reset the counter because a new loop, process, or model turn started. The next step should be a fallback, degraded response, queue for later handling, or human escalation, depending on the business consequence.
Free tools Windows power users keep installed
One-click scans. No signup required.
What durable recovery does
Recovery continues a durable attempt
Durable recovery is for work that must outlive the process executing it. The workflow records completed stages, references to outputs, and enough input or configuration to re-enter safely. If a worker disappears, a new worker loads that state and determines what remains. It does not pretend that local variables, an in-memory call stack, open sockets, or an unfinished stream were preserved.
“Recovery also differs from retry. A retry handles a failure reported by the running handler and can consume retry budget. Recovery continues the same durable attempt after its process disappears.”
Microsoft Learn, Resilience for long-running Microsoft Foundry hosted agents
A checkpoint is useful only if the restarted handler can trust its meaning. A record such as stage=3, status=complete should identify the workflow instance, the exact stage version, relevant input or output references, and the time or sequence at which completion was committed. A watermark can serve the same purpose for streams or batches: it tells the next handler which items have been durably acknowledged.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsRank #2
- 【IP68 Waterproof Multifunctinal Smart Watch】Torjalph fitness activity tracker smart watch is IP68 waterproof technology ,you can wear it when swimming,handwashing,it's rain-proof which means that you perform multiple exercise modes outside without worrying about sweat or rain corroding your fitness smart watch,and this watch equipped with a 230 mAh large capacity battery magnetic charging ,and this sports activity tracker watch can be used 5-7 days If fully charged.
- 【Larger 1.83'' Full Screen Watch & Multi-dial Selection】1.83 inch full HD color touch screen smart watch for men / women provides excellent 240*284 resolution image quality ,This Bluetooth smart watch offers a better touch experience and remarkable immersion and ,more than 200 watch faces for you to choose from App"Da Fit ", you can change the dial style according to your mood to wear keeping the world on your wrist
- 【Accurate Activity Tracker Smart Watch 】sports watch activity tracker uses sweat and dust resistant breathable silicone strap, our Fitness Tracker Smart Watch for Android iOS Phones supports 12 sports modes,walking,running,bicycle,swimming,badminton,basketball,football,rope skipping. This bluetooth calls smart watch for men women is lightweight and stylish, bringing more convenient and relaxing sports experience while also providing more protection.
- 【Your 24 Hour Personal Daily Health Monitor】Fitness watch for women men full touch screen IP67 waterproof bluetooth smartwatch for Android iOS Phones with heart rate blood pressure monitor multifunction Andriod smart watch with steps counter sleep monitor sports fitness watch, monitoring blood oxygen saturation allows you to understand your physical condition in real time to track fitness progress ,to know your health status in time.
- 【More Functions Bluetooth Call Smart Watch】This fitness activity smart watch full screen for iPhone and Android phones includes a bunch of new features including:The bluetooth smart watch can hang up the phone, receive and display WhatsAPP /Text /Facebook /Twitter /Messenger reminder.Support sedentary reminder, alarm clock, stopwatch, timer,weather information,music control, remote camera control compatible with Android 5.0 and above or iOS 9.0 and above smartphones.
Recovery is not deterministic replay
Re-entering a handler from its beginning is common. The handler reads the checkpoint, skips work known to be complete, and executes the next stage. Any operation that ran after the last checkpoint may have an uncertain outcome, so recovery must reconcile it just as a retry would. Persisting a checkpoint does not make an external side effect idempotent, and it does not restore hidden state held only in memory.
Six questions that decide between retry and recovery
1. What kind of failure occurred?
Start with the failure category, not the desired result. A transient timeout from a running request points toward a retry. A process disappearance points toward recovery. An invalid request points toward correction. If the category is unknown, pause and inspect the operation’s outcome rather than guessing.
2. Is replay safe?
Replay safety means the provider and your application tolerate the same call more than once. A read-only lookup is often easier to replay than “send this message,” “charge this account,” or “create this record.” Even a read can be unsafe if it advances a cursor or consumes a one-time token. Document the replay contract for every tool and model operation that a policy may repeat.
3. How long and how interruptible is the work?
A single inexpensive request can often be rerun from the beginning. A multi-stage research, coding, data-processing, or approval workflow may have minutes or hours of useful progress. The longer the work and the greater the chance of deployment or worker interruption, the stronger the case for stage boundaries and durable checkpoints.
4. What progress is worth preserving?
Save progress when recomputing completed stages is expensive, nondeterministic, rate-limited, or produces side effects. If rerunning the whole task is cheap and harmless, a checkpoint may add complexity without much value. Make this a business decision as well as an engineering one: identify which outputs are costly or impossible to recreate.
5. Which external side effects can be observed?
For every tool that changes state outside the agent, define how to discover whether the change happened. Use an idempotency key where the provider supports one, or create a durable operation record keyed by your workflow and action identifiers. Store a result reference, not merely a local boolean. If the provider offers a status query, make that query part of recovery.
6. What happens after exhaustion?
Choose the post-failure path before writing the retry loop. A user-facing agent might return a partial result and explain what remains. A data pipeline might place the item in a review queue. A safety-sensitive action may require a person. Tracing should identify the workflow, stage, attempt number, dependency, and last durable checkpoint so the fallback is actionable.
Rank #3
- [Low Noise High Fidelity] New bee guitar cord with 22AWG oxygen-free-copper center conductors provides enhanced signal transfer, the PVC insulation material and the high-density stranded copper double shield provide noise rejection for quiet operation
- [Ideal for Choice] New bee 1/4 straight and right angle metal connectors, ensures a secure connection in any instrument, pedal, or amp jack. L shape great fro on stage guitar stands
- [Wide Compatible] Suitable for mixer, power amplifier, microphone, effector, decoder, equalizer, electoronic organ, electric guitar, electric bass and other electronic equipments. Widely used in the installation of smart home, karaoke speaker, professional stage audio, etc
- [Easy Disassemble] Making incidental repairs quick and easy! Very nice, flexible cable
- [Dont Hesite to Order] 24 hours customer services and professional technology team are standing by
A combined design for long-running agents
Long-running agents generally combine local retries with durable progress. The following sequence keeps the two responsibilities separate.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute- Decompose the workflow into recoverable stages. Give each stage a clear input, output, completion condition, and owner for side effects. Avoid one giant handler that mixes planning, tool calls, and final publication.
- Persist an initial workflow record. Include a unique workflow ID, input references, configuration or prompt version, current stage, and status. Keep large artifacts in durable storage and retain references in the record.
- Run eligible operations with a local retry policy. Classify the error, check whether the operation may already have completed, apply backoff and jitter, and honor the overall deadline.
- Reconcile uncertain outcomes. When a timeout follows a tool invocation, query the provider or your operation table before issuing the same command. Mark the action completed only after an authoritative result is available.
- Commit a checkpoint at the stage boundary. Write the output reference, stage version, side-effect identifiers, and next stage in one durable transaction where possible. A checkpoint written before the side effect is confirmed can cause unsafe duplication; one written after confirmation may require the side effect to be idempotent during recovery.
- Restart from the checkpoint. A new handler loads the record, verifies that the stage definition is compatible, skips committed work, and begins the first incomplete stage. It should be safe to run this initialization more than once.
- Escalate persistent failures. After the retry budget, mark the stage failed with diagnostic context and invoke the selected fallback. Do not keep the workflow in a hidden loop.
A minimal control flow can look like this:
while workflow.has_incomplete_stage():
stage = workflow.next_stage()
result = reconcile_existing_operation(stage)
if result is None:
result = run_with_bounded_retry(stage.operation,
deadline=workflow.deadline,
honor_retry_after=True)
if result.is_permanent_failure:
return route_to_fallback(workflow, stage, result)
if result.is_unknown:
return wait_for_reconciliation(workflow, stage, result)
workflow.commit_checkpoint(stage, result)
return workflow.complete()The code is illustrative: the durable commit, operation identity, and transaction semantics must match the storage and tools in your system.
How to implement a safe retry policy
Define an allowlist, not a catch-all
Represent retryable conditions explicitly. A generic exception handler that retries every error can resubmit invalid input, conceal authorization problems, and overload a failing dependency. Include the dependency’s documented status codes or error types, but still inspect whether the operation began and whether an external action is observable.
Set both attempt and time boundaries
An attempt limit protects a dependency from repeated calls; a deadline protects the user and the workflow from unbounded latency. Propagate the remaining deadline to nested model and tool calls. If a retry would exceed it, stop and use the fallback path instead of starting work that cannot finish.
Use exponential backoff with jitter
Increase the delay after each eligible failure and randomize it within a controlled range. Cap the delay so one attempt does not wait longer than the workflow can tolerate. If the server supplies Retry-After, use that instruction when it is valid for the operation and still fits the deadline.
Recommended Free Tools
Stop when the situation changes
A changed error is evidence that the original retry assumption no longer holds. Stop when the response becomes a validation error, permission error, or other non-retryable class. Also stop when an attempt returns a definitive result, even if a later client step fails; use reconciliation rather than blindly rerunning.
Account for streaming and local side effects
Model retries are not automatically safe after a stream has started emitting output or when the request includes local side effects. The OpenAI Agents SDK Python documentation describes runner-managed model retries as opt-in, documents retry policies and backoff, and includes protections against replay in cases such as streamed runs after output begins or requests with local side effects unless replay safety has been established: OpenAI Agents SDK Python model documentation. Treat those behaviors as SDK-specific, not as a guarantee shared by every agent framework.
Rank #4
- Digital Family Calendar: BSIMB digital calendar wall touch screen puts an end to your chaotic life. Designed for busy families, the digital calendar keeps you organised and in step with your family. Assign different colours and tasks to each family member to make family scheduling easier and more intuitive.
- Easy Setup and Auto-Sync: Plug in, connect to WLAN and auto-syn your electric calendar in minutes. Digital wall calendar compatible with Google calendar, iCloud calendar, Outlook calendar, Cozi calendar and Yahoo calendar. Add events directly on the device or through the ecalendar app. Electronic calendar stays current whether you're at work or traveling—no manual updates needed.
- All-In-One Home Assistant: Smart calendar builds habits and encourage independence with personalised chores. Display daily meal plan to reduce the daily ‘What's for dinner?’ stress. To-do lists keep track of everything you want. Turning chores into valuable achievements with stars and rewards, making life easier and more enjoyable.
- Manage Anytime Anywhere: BSIMB electronic calendar for family stay connected with the powerful ecalendar app. Access and manage your digital wall calendar anytime, anywhere with free features like stars & rewards, photo screensaver, chore charts, meals plans, to-do lists etc. Electronic planner paid features like Magic Import tailor advanced tools for users with special needs.
- Stay in Touch with Loved Ones: Share photos or videos from your phone to electronic wall calendar with app, instantly update your life moment, and keep in touch with your family, especially those who are far away; you can also surprise your smart calendar digital wall by sending him a wish card; it's also great for keeping tech-illiterate elders in touch with the youngsters.
How to design checkpoints and recovery state
What to store
- Workflow and stage identifiers, including a version of the stage code or schema.
- Input references and immutable configuration needed to reproduce the decision.
- Output references, hashes, or durable artifact locations rather than volatile in-memory objects.
- Completion status, timestamps, attempt history, and the next stage to execute.
- External operation IDs, idempotency keys, and reconciliation status for every side effect.
- Error classification and a human-readable diagnostic trail for fallback or review.
Choose storage for the deployment boundary
Checkpoint storage is a framework and deployment decision. Microsoft’s Agent Framework workflow documentation lists in-memory storage for short-lived work, local file storage for single-machine workflows, and Cosmos DB storage for production distributed, cross-process workflows. These choices are framework-specific; verify current package versions, supported APIs, consistency behavior, encryption, retention, and deployment requirements before adopting one: Microsoft Agent Framework workflow checkpoints.
| Storage option | Suitable boundary | Limitation to account for |
|---|---|---|
| In-memory | Short-lived work where losing progress is acceptable | Process loss erases the checkpoint. |
| Local files | A single machine or local development workflow | Other workers and machines may not see the same files; handle locking and corruption. |
| Distributed durable store | Production workflows that can move across processes or hosts | Plan consistency, access control, schema migration, retention, and operational cost. |
Make checkpoint writes atomic enough
Use a transaction or compare-and-swap condition where the store supports it. A worker should not overwrite a newer stage completion with an older retry. Include a monotonically increasing version or lease owner, and make recovery tolerate a worker that dies immediately before or after a write. Test both edges deliberately.
External side effects: the hardest boundary
Suppose an agent calls a notification service and then times out while waiting for the response. Retrying may send two notifications; refusing to retry may send none. The correct action is to query by an idempotency key or operation ID, then record the authoritative result. If the provider has no query or deduplication facility, create a durable intent record and design a reconciliation process that can compare the intended action with observable state.
For database writes, use a unique business key or an idempotent upsert where appropriate. For file or object creation, derive a stable name and verify its contents before replacing it. For queues, carry a message ID and make the consumer deduplicate it. For a multi-step external transaction, record each confirmed action and its compensation or manual remediation path. A checkpoint alone cannot undo a duplicate side effect.
Observability and failure handling
Every attempt and recovery should emit structured telemetry containing the workflow ID, stage, dependency, operation ID, attempt number, retry classification, delay, deadline remaining, checkpoint version, and final disposition. Trace the boundary where a process disappeared separately from an error returned by a running handler. This lets operators distinguish a slow dependency from a worker crash and a duplicate request from a lost response.
A useful failure record answers four questions: what the agent was trying to do, what it knows completed, what remains uncertain, and who or what owns the next action. Preserve the original error and the reconciliation result; replacing both with “retry failed” removes the information a fallback needs.
Framework-specific cautions
OpenAI Agents API and SDK
The Agents API guidance treats request errors and failures after work has begun differently. Check completed actions before retrying, keep retries bounded, honor server timing instructions, and stop when the error class changes or the limit is reached. In the Python SDK, runner-managed model retries are opt-in and include policy and backoff controls; streaming and local side-effect cases require an established replay contract. Configure these mechanisms deliberately rather than assuming every model call is automatically retried.
Microsoft Foundry long-running resilience
Microsoft’s long-running Foundry agent resilience feature is marked preview. Preview features do not carry a service-level agreement and are not recommended for production workloads. Its documented behavior is still useful conceptually: recovery re-enters a handler from the beginning and relies on durable checkpoints or watermarks, while retry handles a failure reported by a handler that is still running. Recheck availability and behavior before relying on this feature in a deployment.
AWS Agentic AI Lens
AWS recommends decomposing workflows into recoverable stages, classifying failures before choosing a response, retrying transient errors, and using fallbacks for persistent failures. It specifically warns against uniform retry policies, fixed intervals without jitter, retry-only recovery, and insufficient tracing. Those warnings apply even when an agent framework supplies a default retry setting.
Common anti-patterns and fixes
| Anti-pattern | Why it fails | Better design |
|---|---|---|
| Retry every exception three times | Permanent errors waste time and can multiply harmful actions. | Allowlist retryable classes and stop when classification changes. |
| Restart the entire workflow after every crash | Completed stages are recomputed and side effects may duplicate. | Persist stage checkpoints and reconcile work after the last durable commit. |
| Write “done” before confirming the external action | Recovery can skip an action that never happened. | Record intent, confirm the provider result, then commit completion. |
| Assume a timeout means no server action occurred | The request may have succeeded while the response was lost. | Query by operation ID or idempotency key before repeating. |
| Use one global retry policy for every tool | Reads, writes, streams, and one-time actions have different replay risks. | Give each operation a documented replay and failure policy. |
| Hide exhausted retries in logs | Operators cannot tell whether the issue is transient, permanent, or a process loss. | Trace attempts, checkpoints, classifications, and fallback decisions. |
Worked decisions
Short retrieval call times out
The call is read-only, the provider reports a temporary timeout, no result has been observed, and the remaining deadline allows another attempt. Retry with backoff and jitter. If the provider later reports a definitive response, stop; do not run an additional attempt merely because the client missed a notification.
A report-generation workflow loses its worker
The workflow has completed data collection and outline stages, each recorded durably. A new worker loads the checkpoint, verifies the artifact references, and starts the first incomplete stage. It does not spend retry budget on stages that already committed successfully.
A message tool times out after submission
The outcome is unknown and the side effect is not assumed safe to replay. Query the message service with the operation ID or idempotency key. If the message exists, record completion; if it does not and the provider guarantees the query is authoritative, submit once; otherwise route the case to reconciliation or a person.
A tool rejects an argument
Do not send the same malformed call again. Correct the schema, ask the model for a valid value, or return a clear failure. If correction itself is impossible within the workflow’s policy, use the fallback path and preserve the validation error.
Testing the boundary before production
- Inject timeouts before a request is sent, while it is in flight, and after the provider has accepted it.
- Kill a worker before and after every checkpoint write, then verify that recovery neither loses a completed stage nor repeats an unsafe side effect.
- Exhaust the retry budget and confirm that the fallback receives the stage, operation ID, error class, and last known progress.
- Return a different error on a later attempt and verify that automatic retries stop when classification changes.
- Test concurrent recovery workers to ensure stale checkpoints cannot overwrite newer progress.
- Validate retention, access control, encryption, and cleanup for checkpoint and operation records.
There is no universal retry count, checkpoint interval, or performance percentage that determines the right design. The correct values depend on dependency contracts, side-effect risk, workflow duration, and the cost of recomputation. Treat SDK examples as implementation settings to evaluate, not as empirical guarantees.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →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.
