What is a workflow? Traceable tasks built from steps and conditions
Explain predefined control flow and model decisions through inputs, branches, recovery and final states.
Understand the concept
A workflow organizes steps, conditions and dependencies using people, code or models. It need not use AI or eliminate judgment. Workflows typically define control more explicitly than dynamic agents, and both can combine. Validate the whole task and failures, not one successful step.
Control flow can include model and human steps
Anthropic distinguishes predefined orchestration from more autonomous agents and Power Automate organizes triggers and actions. This is a useful control distinction, not universal naming. Models can select branches and humans confirm results. Define step inputs, outputs and missing-data paths rather than guessing to reach completion. High automation is not validated business judgment.
Test normal and failure paths for the same task
For a public-form report, verify a usable file and defined missing-field state. Simulate timeout and recovery without duplicate writes. Keep actual step results and final state rather than model completion claims. No workflow was built or run here. Match input and output versions to avoid stale-file mixtures.
Maintain dependencies and recovery settings
Check dependent steps after updates and retain recoverable versions and cases. Include setup, running and repair effort. Fixed workflows still face format and interface changes. Bound added model decisions and test them rather than rewriting everything for an agent label. Maintain traceable failures, targeted reruns and final verification beyond diagrams.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.