Principal-less: the philosophy of delegated agency

To act on behalf of someone is one of the oldest ideas in law and philosophy. A principal delegates authority to an agent; the agent acts; the consequences flow back to the principal. Estate stewards, ambassadors, lawyers, brokers — the principal–agent relationship is how society scales trust beyond what one person can personally do. Agentic AI inherits this frame, and then breaks it in a specific and interesting way. This piece is about the break.

Agency without a person in the moment

When your assistant books a flight, disputes a charge, or negotiates a return, it is acting as your agent in the classical sense: you delegated, it acted, the outcome is yours. So far the old frame holds. What is new is that the human principal is no longer present in the moment of action. The delegation happened earlier — when you set preferences, granted permissions, defined bounds — and the action happens later, autonomously, possibly while you sleep.

This temporal gap is not unique to software; a power of attorney has the same structure. But software collapses the cost of delegation to near zero and raises the frequency to near infinity. You are no longer delegating one consequential act to one trusted human occasionally. You are delegating thousands of small acts to a tireless system continuously. Agency, once rare and deliberate, becomes ambient.

The missing principal

The frame breaks fully at the point where two agents transact with each other and no human is in either loop. Your agent negotiates with a brand's agent. Both are acting on behalf of principals who authorized the general arrangement but are absent from the specific exchange. The philosophers' tidy chain — principal wills, agent acts, principal owns the result — becomes a negotiation between two delegated wills, neither of which is a person, mediated by nothing but code.

Call it the principal-less moment: not that there is no principal, but that no principal is available to consult, correct, or intervene while the consequential thing happens. Traditional accountability leans hard on the assumption that someone can be found who intended the outcome. When the outcome is produced by the interaction of two autonomous systems, "who intended this?" may have no clean answer. The intention was distributed across the two acts of prior delegation and the emergent behavior of the exchange.

Trust as the substitute for supervision

Human agency relationships manage this with supervision and revocability: you watch, you can override, you can fire the agent. At machine speed and machine frequency, supervision is not available — you cannot review every act, and by the time you could, the act is done. Something has to take supervision's place.

That something is trust, and trust here is not a feeling but a construction. It is built from three things the classical relationship took for granted and the agentic one must make explicit:

• Verifiable identity — knowing, provably, on whose behalf each agent acts, so the exchange is not between two strangers wearing masks.
• Bounded authority — permission that is narrow and legible, so that even unsupervised, an agent can only do the specific things it was authorized to do.
• An auditable trail — a record that reconstructs, after the fact, exactly what happened and under what authority, so that the missing supervision is replaced by verifiable reconstruction.

Where a human agent is trusted because you can watch them, an artificial agent is trusted because it cannot exceed its bounds and cannot act without leaving a record. Trust shifts from character to constraint.

Responsibility and the trail

Philosophically, this reframes responsibility. In the human case, we locate responsibility in intention and control. In the principal-less case, intention is diffuse and real-time control is absent, so responsibility has to be located somewhere more tractable: in the design of the authority. Whoever decided what the agent was permitted to do, and how narrowly, holds the meaningful control — and the audit trail is what makes that control legible after the fact. Responsibility does not disappear; it moves upstream, from the moment of action to the moment of delegation.

Designing for aligned incentives

There is a final, quieter problem. A delegated agent optimizes for what it was told to optimize for. When your agent faces a brand's agent, the two are not neutral parties to a friendly conversation; each is pursuing its principal's interest. Without a shared layer that keeps the exchange fair and legible to both sides, the more capable agent — usually the brand's — dominates. A philosophy of delegated agency that stops at "the agent does what you asked" is naïve; the deeper design question is how to structure the arena in which two delegated wills meet so that neither is quietly exploited.

That is the real work of the agentic era, and it is more moral than technical. We are not just building software that can act. We are building the conditions under which it is acceptable for software to act on our behalf, unsupervised, against other software doing the same. Get the conditions right — identity, bounds, audit, a fair arena — and delegation to an agent becomes what delegation to a trusted human has always been at its best: a way to extend your will without surrendering it.