Build a durable workflow for a multi-day business process

A process that waits days for approval cannot depend on a live request, a model conversation or one worker’s memory. Durability comes from explicit state, persisted history and repeatable transitions. The model may propose content; the workflow engine owns progress.

Build the decision artifact

Name every state, permitted transition, timeout and terminal outcome. Persist intent before an external effect, assign stable workflow and idempotency identifiers, and checkpoint after observable transitions. Separate retryable technical failure from a business exception that needs an owner. Version state and event schemas before deployments overlap.

The enterprise delivery lab makes this reasoning executable with synthetic data:

from enterprise_cases import durable_workflow_case

result = durable_workflow_case()
assert result["state"] == "exported"
assert result["checkpoint_count"] == 5
assert result["resumable"] is True
assert result["in_memory_only"] is False

The synthetic trace spans three days and reaches exported through five persisted events. The result can be rebuilt from history, so a replacement worker does not need hidden context. In a fuller design, each transition also records actor, command, result and correlation id.

Test the failure path

A durable store does not make unsafe replay correct. Re-executing an external write can duplicate it; skipping a completed write can lose it. Couple checkpoints with idempotent downstream contracts or reconciliation. Test timeouts and worker termination between the external effect and local acknowledgement.

Save evidence that another reviewer can inspect:

• state and transition table
• event schema with versioning policy
• kill-and-resume test
• idempotency and reconciliation record

This exercise aligns with the enterprise delivery, secure integration, operations and leadership work in the FDE for Professionals course. The course link describes the pathway; this local draft and its synthetic verifier do not claim that a client deployment, security approval or production result occurred.

Practice task

Draw the states for a three-day approval. Kill the worker after each transition and show which persisted fact lets another worker resume safely.

Continue learning

This article is part of the Enterprise AI delivery and architecture sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Plan an enterprise connector migration with compatibility tests.
• Continue with Design compensating actions for partially completed operations.

Reference: Azure Architecture Center: Scheduler Agent Supervisor.