When to Use an Agent

Use an agent if:

1. Task isn't known in advance — subtasks change with user input
2. External access mandatory — web, DB, APIs
3. Multi-step + adaptive — steps change observation-to-observation
4. Flexibility > speed — user is OK waiting 30s, value is high
5. Outcome verifiable — there's an evaluable result

Don't use an agent if:

• A single prompt solves it → don't
• 1s latency required → too slow
• Error cost very high (financial transfer, medical diagnosis) → workflow + HITL
• Predictable + repeatable → DAG/workflow suffices

Test: "Can I write this as a workflow?" If yes → write it. If no → then agent.