Free tools Windows power users keep installed
One-click scans. No signup required.
It should resume from a durable checkpoint, not depend on one process surviving all 50 steps. The workflow needs enough saved state to continue, a record of which actions actually completed, and safeguards against repeating external side effects. “Step 37” is an illustrative scenario, not a measured failure rate; what happens next depends on the workflow’s checkpoint and retry design.
Retrying is not the same as resuming
A retry attempts an operation again. A resume restores saved progress and state so the workflow can continue from an appropriate boundary. Microsoft Foundry’s documentation treats recovery and retry as distinct ideas; its long-running agent resilience feature is marked as a preview, so availability and behavior should be checked for the specific environment.
As an Amazon Associate I earn from qualifying purchases.
A process restart alone is not recovery. If the system has not persisted the information needed to continue, it may have to reconstruct state or repeat work. Conversely, a checkpoint can let an orchestrator avoid re-running completed work, but it does not guarantee that every outside action is safe to repeat.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What to do when a run stops at step 37
- Find the latest durable checkpoint. Check the workflow’s execution history to establish which saved boundary is available. The checkpoint should contain the state needed for continuation, not merely a step number.
- Establish what happened after that boundary. Inspect recorded inputs, outputs, status transitions, and tool or service responses. Identify any action that may have succeeded externally even if the workflow did not save its response.
- Reconcile uncertain external actions. Before retrying a payment, message, record creation, or other consequential operation, check the destination system or use a deduplication mechanism. A missing response in the workflow is not proof that the external system did nothing.
- Validate the state to be resumed. Confirm that required outputs and downstream inputs are present and valid. Then restore the persisted state and resume from the appropriate boundary.
- Escalate when the outcome is ambiguous or unsafe. If you cannot establish whether an irreversible action occurred, route it to reconciliation or human review rather than blindly replaying it.
Microsoft Agent Framework Workflows documents resuming from a selected checkpoint. Microsoft’s Durable Task extension documents checkpointed agent calls and recovery without re-executing completed calls. Those are framework-specific documented behaviors; they do not make arbitrary external actions exactly-once.
What a useful checkpoint needs to preserve
A checkpoint is useful only if it captures the information needed to continue safely. At meaningful stage boundaries, persist the workflow state along with the inputs and outputs that later stages depend on. Make state transitions explicit—for example, distinguish “request sent,” “response received,” and “result committed” where the external system supports those distinctions.
- Progress: which stage or boundary has completed.
- Continuation state: the context and data required by subsequent stages.
- Stage inputs and outputs: enough detail to validate downstream work and diagnose a bad result.
- External-action status: the available evidence of whether an action was attempted, accepted, or confirmed.
- Execution history: identifiers and events that let operators trace the run across agents, tools, queues, and services.
Checkpoint frequency is a trade-off. Boundaries that are too far apart can mean more work must be reconstructed or repeated after a failure; boundaries that are too fine-grained add state and orchestration overhead. Choose boundaries around meaningful, recoverable stages rather than assuming every individual model or tool call must be a checkpoint.
Make repeated work safe before enabling automatic retries
An idempotent operation produces the same intended result when repeated with the same input, rather than adding another side effect. AWS guidance recommends designing for idempotency because a workflow may retry after an error without knowing whether the original operation completed.
For non-idempotent actions, use a stable operation identifier or destination-supported deduplication where available, check the destination before repeating, or require human approval. These approaches reduce duplicate effects, but their suitability depends on what the external service supports. Orchestration replay and application-level deduplication solve different problems: replay controls workflow execution; deduplication or reconciliation controls effects outside it.
Match the recovery response to the failure
- Transient error: a bounded retry may be appropriate for a failure likely to clear, such as a temporary service or network problem. Use a limit and backoff suited to the operation rather than retrying indefinitely.
- Persistent error: stop repeating the same failing operation and use a defined fallback, such as an alternate route or a repair path, if the workflow supports one.
- Ambiguous or irreversible outcome: reconcile the destination state or ask a person to review it before proceeding.
- Unrecoverable state: preserve the execution history and surface the failure for intervention instead of silently restarting from the beginning.
AWS’s Agentic AI Lens recommends recoverable stages, classifying failures for targeted retries, and end-to-end distributed tracing. That guidance supports treating retry policy, fallback, escalation, and traceability as parts of one recovery design.
How documented approaches differ
These products and frameworks describe different layers of workflow continuity. Their documentation does not establish one universally best choice, and the listed behaviors should not be read as equivalent guarantees.
| Approach | Documented capability or emphasis | Important boundary |
|---|---|---|
| Microsoft Agent Framework Workflows | Workflow checkpoints and resumption from a selected checkpoint. | The documentation describes workflow resumption; it does not establish that every external side effect is safe to replay. |
| Microsoft Durable Task extension | Checkpointed agent calls and recovery without re-executing completed calls. | This behavior is specific to the extension and its orchestration; do not generalize it to arbitrary external actions. |
| AWS guidance and services | Guidance on persisting state, staged recovery, idempotency, and redrive. | Guidance does not establish that a particular service configuration fits a given workload. |
| Temporal | Temporal describes Temporal Cloud on AWS as a managed workflow orchestration service. | The cited product information does not state a comparable checkpoint granularity or a guarantee for external side effects. |
| OpenAI Agents SDK | Documentation describes the agent run loop, including tool calls, handoffs, and strategies for carrying state into later turns. | Application-level agent continuity is not proof of a general-purpose durable 50-step workflow engine. |
Questions to settle before choosing an architecture
- Checkpoint contents and boundaries: what state is saved, and how much work could need recovery between boundaries?
- Resume semantics: after recovery, which completed stages or calls can run again?
- Side-effect protection: which operations are idempotent, and how will ambiguous outcomes be reconciled?
- Failure policy: what errors merit a bounded retry, a fallback, or escalation?
- Observability: can an operator trace one execution across agents, tools, queues, and external services?
- Human gates and operations: which actions need approval, and who owns runtime operations in the selected managed or self-operated setup?
No general statistic about how often long multi-step workflows fail late, or what those failures cost, is established here. The 50-step example is a way to frame the recovery problem, not a forecast.
Recommended Free Tools
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.
