Every agent project I have seen go wrong went wrong at the same step: someone connected a capable model to a real system without first deciding which decisions that model was allowed to make on its own. The model was fine. The permission was missing.
The decision inventory
Before you write a line of orchestration code, list every decision the agent will touch. For each one, answer four questions:
- What does a wrong answer cost? A wrong product recommendation costs a click. A wrong refund costs money. A wrong price costs money and trust.
- Can it be undone? An email cannot be unsent. A price can be reverted. A refund can be reversed within a window if you design for it.
- How often does it happen? A thousand decisions an hour cannot have a human on each one. Ten a day can.
- Who owns it today? If a person owns it now, that person is your evaluator later.
Three tiers of autonomy
| Tier | The agent may | Example |
|---|---|---|
| Suggest | Draft, score, rank. A human acts. | Reply drafts, lead scoring, price suggestions |
| Act within bounds | Act alone inside hard limits set in code | Refunds under a cap, price moves under a step limit |
| Act and report | Act freely, log everything, humans audit samples | Search ranking, internal routing |
Every decision starts in Suggest. It earns its way to the next tier with evidence, which is what lessons 3 and 4 are about.
Do this today: write the decision inventory for one agent you are building or buying. One page. If you cannot fill in the "wrong answer cost" column, you are not ready to automate it.