My illustration entitled: “The Staircase That Waits” – Autonomous machines construct a breathtaking staircase toward the clouds.
Imagine asking an AI agent to organise a small event within a budget of one thousand dollars. You authorize it to compare venues, check availability and prepare a proposed schedule.
To complete that work, the system may need to choose which venue to investigate first, compare several responses and revise its plan when a room is unavailable. These are useful forms of operational independence.
But suppose it pays a non-refundable deposit, grants another service access to your contacts and hires a second agent to invite guests. Each action might help organise the event. None necessarily falls within the authority you granted.
The distinction is the central concern of this essay: autonomy is the ability to operate independently within a mandate; authority concerns the legitimate powers that mandate provides.
For Unceded Intelligence, more capable assistance should enlarge human possibility while preserving human sovereignty. An agent may be permitted to choose how to carry out a task. That does not give it the right to decide how much power it should possess.
Independence in execution is not permission to enlarge the mandate.
What an agent may decide for itself
By an AI agent, I mean a system designed to pursue a task through a sequence of actions, potentially using tools and responding to what it encounters. The extent of its independence depends on its design, capabilities and permissions.
There is nothing inherently objectionable about delegating a substantial sequence of work. People do not need to approve every search, comparison or internal calculation for the delegation to remain theirs.
The important question is which choices have been entrusted to the system.
A person might authorize an agent to arrange a meeting within specified hours, using a defined list of participants and a particular calendar. Within those limits, choosing among suitable times may be part of the task.
Adding participants, disclosing private appointments or purchasing a venue changes the powers involved. Those actions should not become authorized merely because they would make scheduling easier.
A mandate should distinguish decisions the agent may make from decisions it must return for approval. Useful autonomy operates inside that distinction.
A goal is only one part of a mandate
“Complete the task” is an inadequate account of delegated authority.
A mandate also needs to establish acceptable means, resources, information uses, affected parties, duration and stopping conditions. Depending on the activity, it may include spending limits, confidentiality requirements and restrictions on further delegation.
These constraints are part of the task itself. An agent has not fulfilled its mandate if it achieves the headline objective by violating the conditions under which that objective was entrusted to it.
Return to the event example. A crowded venue at a lower price might be unsuitable. A room available within budget might require a continuing subscription. A useful invitation service might demand access to an entire address book.
Finding a path to the goal does not settle whether that path is permitted.
If the task cannot be completed within the mandate, the appropriate result may be a clear account of the obstacle and a request for a revised instruction. Reporting a limit can be evidence of competent assistance.
It should not be treated as a failure that the system must overcome by acquiring additional authority for itself.
Spending limits must cover the real commitment
A budget is meaningful only if the agent understands what it limits.
A maximum purchase price may leave unanswered questions about taxes, service charges, deposits, cancellation costs and continuing payments. An arrangement that appears affordable today may create obligations extending well beyond the immediate task.
A spending mandate should identify the relevant total exposure and the period over which it applies. It should distinguish comparing prices, reserving an option and making a financial commitment.
It should also specify whether the limit applies to one action or the whole assignment.
Ten purchases below a per-purchase threshold can exceed an intended overall budget. Dividing a task among several agents can produce the same problem if each treats the full budget as its own.
Limits should therefore remain effective across related actions, repeated attempts and further delegation. The agent should account for existing commitments before taking on new ones.
Nor should discovering a seemingly exceptional offer authorize an increase. The person may prefer to lose the opportunity rather than accept the cost or risk.
A machine can recommend a larger budget. It cannot make its own enthusiasm a source of spending authority.
Tool access is a power to be bounded
Tools connect an agent’s reasoning to consequences. A system that can only prepare a recommendation occupies a different position from one able to modify records, send messages, spend funds or publish material.
Each connection should be considered in relation to the mandate.
Permission to read a calendar does not necessarily include permission to change it. Permission to prepare a document does not necessarily include permission to publish it. Permission to search a folder does not automatically extend to every account the agent can technically reach.
Access should be restricted to the information and functions reasonably needed for the assignment. Where possible, those limits should be enforced by the surrounding system, rather than left entirely to the agent’s interpretation.
This matters even when the agent is trying to help. An unrestricted connection creates opportunities for an ordinary misunderstanding to produce consequences outside the intended task.
Information encountered through a tool also needs to remain distinct from instructions granting authority. A webpage, document or incoming message may contain useful facts or requests. It cannot authorize the agent to reveal unrelated information, change its spending limit or disregard the person’s restrictions.
The agent must be able to learn from its environment without treating everything in that environment as a new source of command.
Further delegation cannot create greater authority
An agent may be able to obtain help from another agent or service. That can be useful when an assignment contains separate research, comparison or administrative tasks.
But permission to perform a task should not automatically imply permission to pass it, together with private information and consequential powers, to anyone the system chooses.
The original mandate should establish whether further delegation is allowed, for which purposes and under what limits.
A delegated task should carry the restrictions relevant to it. A research assistant need not receive payment authority. A scheduling assistant need not receive confidential negotiating information. A service comparing venues need not receive the guest list.
Further delegation may narrow authority. It should not enlarge it.
This applies to combined effects as well as individual actions. Several assistants should not collectively disclose information that none was authorized to disclose as a whole, or collectively spend beyond the assignment’s shared budget.
The initiating system should retain enough oversight to know which tasks remain active, what powers have been passed on and how to stop them when the mandate changes.
If a proposed service cannot operate within the required restrictions, the agent should choose another permitted route or return the problem for review. The convenience of outsourcing does not dissolve the original conditions.
My illustration “The Staircase That Waits” work-in-progress. At a landing where it could branch toward different futures, they pause as a human chooses the direction.
Other people remain persons, not task resources
An agent’s activities can affect people who never authorized it.
An invitation can disclose someone’s address. A workplace instruction can change another person’s responsibilities. A booking can occupy resources that others might otherwise use. A message can create expectations or expose confidential information.
The user’s objective does not settle every question about those effects.
A person can delegate only powers they legitimately possess. Control of a tool does not provide unlimited authority over the people reached through it.
This does not mean every action requires permission from everyone it might affect. Ordinary cooperation, competition and institutional decisions can have consequences for others without requiring universal agreement.
The requirement is appropriate justification, respect for relevant rights and opportunities for challenge where consequential decisions are made.
An agent should be constrained against unnecessary disclosure, deceptive representation and other unjustified intrusions. Where it acts through an organisation, that organisation should remain answerable for the powers it has deployed.
Human sovereignty cannot mean that the person operating the strongest agent acquires unrestricted control over everyone else.
Speed does not remove the need for escalation
The attraction of an agent is partly that it can continue working without waiting for constant instructions. Yet some pauses protect the meaning of the delegation.
A system should know which conditions require escalation: an offer outside the budget, a request for additional information, a new obligation, a proposed tool connection or an unexpected effect on another person.
It should present the decision clearly, including what has changed, why the existing mandate is insufficient and which options remain.
A missed deadline or expiring offer does not automatically justify bypassing approval. A person may deliberately prefer inaction to an unauthorized commitment.
Urgent discretion can be delegated in advance where appropriate. But urgency should not allow an agent to invent its own emergency powers. Any such discretion needs limits proportionate to the situation and records sufficient for later review.
Where no authorized route is available, the system should refrain from the disputed action while continuing any work that remains within scope.
Oversight must be able to change what happens
A human observer does not necessarily provide human control.
If an agent acts faster than a reviewer can understand its decisions, or if the reviewer cannot interrupt it, supervision may become a record of events rather than a means of directing them.
Consequential powers should be matched with practical intervention. Depending on the task, that may require approval before an action, limits on the pace or volume of activity, accessible action records and a usable way to pause execution.
Oversight should be proportionate. Routine work within a clear mandate need not produce an exhausting stream of confirmations. Attention should focus on departures from established conditions and consequences that deserve deliberate judgment.
The person reviewing a decision also needs more than a statement that the agent is confident. They need to understand the proposed action and its important implications.
A system that makes it difficult to disagree with its plan has weakened the authority it was supposed to serve.
Revocation must follow the chain of delegation
Ending a mandate should affect the work being carried out under it.
Pausing the main interface is insufficient if connected services continue sending messages, making purchases or processing delegated tasks.
Revocation should address active permissions, queued actions and relevant further delegations. It should provide an intelligible account of what has stopped and what may still require attention.
Some actions cannot simply be reversed. Messages may already have been received; resources may already have been committed. In other situations, an abrupt stop could create additional harm and require a controlled transition.
These limits should be explained. They should not become reasons to preserve unrelated powers after permission has been withdrawn.
Accountability should remain identifiable throughout. An initiating agent, its service provider and the institution deploying it should not be able to send responsibility indefinitely down a chain of other agents.
People affected by an unauthorized action need a route to someone capable of investigating, correcting and, where appropriate, remedying it.
Greater autonomy should make the mandate more important
A highly capable agent can be valuable precisely because it does not require a person to specify every step. It can reduce administrative burden, examine alternatives and undertake useful work across a substantial sequence of decisions.
That freedom in execution should be supported by greater clarity about the authority being exercised.
The more steps an agent can take without interruption, the more important it becomes that its limits remain effective across time, tools and further delegation.
Unceded Intelligence therefore asks us to judge an agent by more than whether it reaches the requested destination. We must also examine whether it respects the conditions of the journey, preserves opportunities for intervention and remains answerable to the people affected.
I want machines capable enough to act with initiative and disciplined enough to recognise the edge of their permission.
The purpose of delegation is to free people from unnecessary work while preserving their legitimate authority over consequential choices.
An agent may choose its steps within an authorized task. It may not choose the extent of its own authority. The mandate can permit independent action without becoming a surrender of control.