W4-011 · Condition 04 — The agent workforce

Scoped authority

Scoped authority is a bounded, explicit, revocable statement of what an agent may do, issued by an accountable principal and verifiable by a counterparty who does not trust the agent.

Published 2026-08-20UnrevisedAlso: bounded authority

What an agent cannot do, it cannot do at any speed.

The instinctive control for autonomous software is monitoring: watch it and intervene. This fails on arithmetic. A fleet produces actions faster than any staffable review capacity, so monitoring degrades into sampling, and sampling catches the average case rather than the damaging one.

Scoped authority is the structural alternative. Authority is stated in advance, denied by default, expressed in terms a counterparty can check, and withdrawable at a speed that matters. It moves control from detection to prevention.

For a founder the practical consequence is that authority design is product design, not a compliance task to schedule later. A system built on inherited credentials cannot have scopes retrofitted, because there is nothing to attach them to.

Related entries

Delegation chainA delegation chain is the traceable sequence of grants from an accountable legal person through every agent that acted on their behalf — the structure that makes autonomous action attributable to someone who can answer for it.Audit trailAn audit trail is a system-written record of what was instructed, by whom, under what authority, and what resulted — sufficient for a party outside engineering to reconstruct any action after the fact.Agent-native companyAn agent-native company is one whose operating work is performed by a persistent fleet of AI agents under human governance from the outset, rather than by staff whose tasks were later automated.