A strong prompt in chat is not automatically an agent. Promotion makes sense when the work is repeatable, tool-backed, and auditable.
Decision checklist
| Question | Prompt enough? | Agent candidate? |
|---|---|---|
| Same steps weekly with same sources | Often yes | If integrated |
| Must call APIs or CRM | No | Yes |
| Needs state across hours/days | Rarely | Yes |
| High-risk external actions | Human send | Agent draft + gate |
Promotion path
- Freeze prompt version that passes eval.
- List allowed tools and data (boundaries).
- Add logging and human send gate.
- Run pilot with override metrics before removing human from loop.