Workflow or agent: choose the control you need

Learning outcome: Identify when a fixed sequence is sufficient.

A workflow follows a defined sequence or set of branches. An agent uses model decisions to select some actions as work unfolds. The distinction is about control, not a marketing label. Both can call models and tools, and both need error handling.

Start with the required outcome and allowed actions. If the process is predictable, an explicit workflow may be easier to inspect. If the next action depends on uncertain information, bounded agent decisions may help. Set stopping conditions, time or call budgets and a route to a person. More autonomy is useful only when the extra flexibility outweighs the additional evaluation burden.

Worked example

A fictional invoice intake process always extracts fields, validates a reference and sends a draft to a reviewer. It does not need an agent to invent its own sequence. An investigation task may need to choose which approved source to inspect next, but should still stop when the evidence is unavailable or its search budget is exhausted.

Try it yourself

Compare those two tasks. Mark fixed steps, model-selected steps and actions requiring approval. Give each process an explicit stopping rule.

You can do this exercise on paper. If using an assistant, keep to the fictional material and an approved tool.

Check your understanding

What is a useful reason to add agent choice?

1. The next permitted step depends on information discovered during the task
2. The word agent sounds more advanced
3. Every workflow must be autonomous to use AI

Workbook

Download the course workbook to record your answer and compare it after the knowledge check.

Further reading

• Anthropic: building effective agents
• Anthropic: writing tools
• MCP architecture