Autonomy is a contract, not a setting. The contract says what an agency is allowed to decide on its own, what it has to send back to the operator, and what the consequences are if it gets either wrong. That contract is the Decision authority framework, and it is the part of the stack that most operators under-build first and regret later.
This essay is the framework. Five fields per agency, applied to every routine decision the agency will make.
Field one: the granted authority.
The granted authority is what the agency is permitted to do without checking in. For the Outbound Operator, that might be: research prospects from a saved list, draft a personalized message, send the message inside a daily send cap. The grant is concrete. It is written down. It is referenced inside the agency's prompt and inside the audit log.
The granted authority should be the smallest set that still produces useful work without operator involvement. Start narrow. Widen as the audit log earns trust.
Field two: the bounds.
The bounds say where the granted authority stops. The outbound agency may send to prospects on the saved list. It may not send to prospects outside the list. It may use the standard message template. It may not write custom legal commitments. It may schedule meetings inside the operator's preferred slots. It may not commit pricing.
The bounds are the safety rail. They are what makes a wider grant tolerable. The wider the grant, the more explicit the bounds have to be.
Field three: the escalation triggers.
An escalation trigger is the condition that takes a decision out of the agency's hands and sends it to the operator. For the outbound agency: a reply that asks for pricing escalates. A reply that asks for the operator's calendar at a non-standard time escalates. A prospect that mentions a regulatory or compliance term escalates.
Triggers are written down, not learned. Learning to escalate is too slow and too lossy. Triggers are explicit so the agency does not have to be clever about whether to ask.
Field four: the escalation interface.
The escalation interface is the format in which the agency surfaces a decision the operator has to make. The minimum is: what is the decision, what are the options, what is the recommended option and why, what is the deadline. A good escalation is a one-glance approve or redirect.
Without an interface standard, every agency develops its own. The operator becomes the translator across formats. The escalation interface is enforced by the Chief of Staff so every agency surfaces decisions the same way.
Field five: the audit pattern.
The audit pattern is how every decision the agency made gets recorded so the operator can review later. The minimum is: timestamp, agency, decision, inputs that informed the decision, outcome. The audit log becomes the trust-building artifact between the operator and the agency.
The audit log is also the input for widening the grant. If the agency has made a thousand routine decisions inside the bounds and the outcomes look good, the operator can widen the grant or relax a bound. If a class of decisions is consistently producing bad outcomes, the operator narrows the grant or tightens a trigger.
How the framework matures.
A new agency starts with a narrow grant and a wide set of escalation triggers. The audit log fills up. The operator reviews. Triggers that fire on outcomes the operator would have approved get retired. Grants get widened. Bounds get loosened where the agency has earned it. The agency becomes more autonomous over time the same way a junior hire becomes a senior one.
This is the part of the stack that turns AI agents into real operating leverage. Without it, autonomy is either too restrictive to be useful or too permissive to be safe. With it, the operator can confidently widen the agency's scope quarter over quarter.
Six production-ready Blueprints. Each one extracted from a working business. Drop into your AI environment and start operating.
Browse Blueprints →