Autonomy budget

A set of measurable limits on an agent's reach, such as allowed tools, data scopes, time, steps, spend, destinations, and maximum effect before the system stops or changes route.

An autonomy budget answers how much reach an agent has for a particular job. A support agent might read one customer's order, issue refunds up to a tested amount, and route larger or unusual cases to another control. A research agent might have broad read access but no send or delete capability. The budget can be enforced with permissions, counters, transaction limits, sandboxes, approval gates, or automatic shutdown conditions; it does not depend on a person watching every run.

Builder example

Explicit limits contain the effect of a bad plan, a misleading source, or a tool error. They also prevent the opposite failure, where routine reversible work waits behind unnecessary approval. The useful unit is the action and its consequence: a $10 test purchase, a message to an internal test address, and a production refund need different limits even if the same model proposes them.

An agent can compare vendors, request quotes, and recommend a supplier.

Give research tools broad read access, cap transaction size and approved vendors, and route any limit crossing to the organization's authorization path.

Common confusion: An autonomy budget is not a single scale from manual to automatic. An agent can have wide latitude to read and compare records, a narrow transactional limit, and no permission to delete data. Crossing a limit can stop the run, request approval, invoke a specialist, switch to a simulation, or require a new authorization.