A useful working distinction
Anthropic distinguishes predefined workflows from systems where a model directs its process and tool use. Products use the word agent differently, so inspect what the system can actually do. A process can contain model-generated text without allowing the model to choose external actions.
Compare the operating model
| Dimension | Workflow | Agent |
|---|---|---|
| Path | Predetermined sequence or routing | Model can choose subsequent steps |
| Good starting task | Known inputs and acceptance rules | Variable task requiring investigation |
| Review focus | Each step and its output | Tool choices, progress and final outcome |
| Control needs | Validation and exception handling | Permissions, stopping rules and action checks |
Illustrative marketing tasks
A weekly report that reads specified metrics and produces a fixed summary is a strong workflow candidate. Investigating an unfamiliar reporting gap may need flexible evidence gathering. Neither example justifies allowing automatic budget changes. Drafting and acting are separate permission decisions.
Run a bounded comparison
Test both approaches on representative and difficult examples. Measure correct completion, human correction time, total latency, cost and failure recovery. Keep the tool access equivalent so the comparison is meaningful. These are our recommended evaluation criteria.
Choose the smallest sufficient system
A single model call or a fixed sequence may perform adequately. Add autonomy only when it improves a defined outcome enough to justify its costs and controls. Re-evaluate when models or task requirements change.
Choose the smallest system that handles the task
Consider a weekly report that uses fixed sources and a stable format. A defined workflow may be easier to test and maintain than an agent that decides which steps to take. If a research task requires choosing new sources and resolving unexpected gaps, some agent behaviour may be useful, provided its tools and review points are bounded. The choice follows the task, not the appeal of the label.
Evaluate both options on representative examples and failures. Compare accuracy, traceability, operating cost and the effort required when something goes wrong. A system that produces a convincing output while hiding its sources may be unsuitable for an evidence-sensitive business decision. Keep a human owner for exceptions and material actions. Expand autonomy only when the observed performance supports it and the business has approved the resulting scope.
- Use a workflow when the sequence and decision rules can be defined clearly.
- Consider an agent where bounded judgement is necessary within the task.
- Test missing data, tool errors and unsuitable instructions before routine use.
Frequently asked questions
When is a predefined workflow preferable to an agent?
Use a workflow when the required steps are known and repeatable. Consider an agent when choosing the steps is part of the task. Compare correct completion, control and cost before choosing the more flexible architecture.
When is a fixed workflow preferable?
When the task and decision path are stable, a workflow can be easier to test and maintain. Use additional autonomy only where it provides a useful benefit.
How do we compare the two approaches?
Test the same representative tasks and assess complete outcomes, errors, review effort and cost. Choose the smallest system that meets the requirement.
Sources and further reading
Primary material for the concepts discussed here. Our practical checklists and illustrative examples are editorial recommendations. Source links open in a new tab.