Map the work before naming the technology
Start with the outcome, owner, trigger, inputs, rules, judgment points, systems, exceptions, and proof of success. The word agent should arrive after the workflow is understood, not before.
Choose the smallest useful intervention
A prompt helps with a one-off output. An assistant supports a bounded domain while the user remains in control. Automation executes known steps. An agent chooses among actions inside explicit boundaries. More autonomy creates more requirements, not automatically more value.
Separate proposal from execution
For consequential work, let the system prepare, classify, compare, or recommend before it acts. Structured outputs, validation, approval, idempotency, budgets, and recovery paths make autonomy governable.