Choosing autonomy levels for each role
Autonomy is not a single dial. It is a per-role decision shaped by reversibility, blast radius, and how confident you are in the work product.
When operators ask 'how autonomous should my agents be?' the honest answer is: it depends on the role. A copy agent that drafts blog posts can run on a long leash. An ops agent that touches billing should not.
Three useful tiers
- Suggest only — agent proposes, human ships every output.
- Approve before send — agent drafts and queues; a human signs off.
- Run autonomously — agent executes; human reviews exceptions.
Most teams over-rotate to tier two for everything because it feels safe. The cost is hidden: reviewing 200 drafts is the same drudgery you were trying to escape. Reserve approval for irreversible or customer-visible actions and let the rest run.
Reversibility is the key heuristic
Drafting an email = reversible. Sending it = harder. Refunding a customer = even harder. Set autonomy to match the cost of being wrong, not the convenience of the operator.