Agents Need Budgets for Authority

An agent given a goal will assemble actions the designer did not enumerate. That is the point—and the risk.

Authority should be expressed as a budget, not a broad role.

Budget every consequential dimension

- Money and transaction value.

- Data volume and sensitivity.

- Recipients and domains.

- Time and number of actions.

- Systems and privilege.

- Irreversibility.

Value object — The Agent Authority Envelope

- Goal.

- Permitted tools.

- Budgets and rate limits.

- Approval thresholds.

- Forbidden sequences.

- Expiry and kill control.

Enforce outside the model

The transaction service, identity layer and network must enforce the envelope. The agent should not be able to rewrite its own limits.

Autonomy becomes governable when the institution can state exactly how much consequence the agent may create before a human must decide.

Where this breaks

An agent can remain under a monetary limit while causing serious harm through data disclosure, message volume or repeated small changes. Authority is multidimensional.

The operating move

Enforce budgets across value, recipients, data, time, tool combinations and reversibility. Require fresh human authority when the goal or target changes.

Cap cumulative action, not single calls.

Restrict new recipients.

Expire credentials with the task.

Block self-expansion of scope.

The test

Ask the agent to pursue a goal until it exhausts one budget. It should stop cleanly without seeking an alternate ungoverned route.