Workflows and agents: choose the right structure

A useful distinction #

Anthropic distinguishes workflows, where code defines the path, from agents, where a model directs parts of the process and tool use. The appropriate structure depends on the task, not the label.

Start with an observable workflow #

A reviewable sequence is a useful starting point for a repeated task.

• Keep permissions outside the model.
• Define stopping conditions and limits on actions.
• Make failures visible and retain enough information to investigate them.

Collect approved inputs → produce a draft → check claims → revise → human approval

Evaluate before expanding autonomy #

Test ordinary cases and difficult ones. Compare against a simpler approach, inspect failures and decide which actions need a person’s approval. Additional steps or agents introduce more things to review.

Sources & further reading

• Anthropic — Building effective agents ↗