The terms your agent accepted

Nobody reads the fine print. This has been true for so long that it has stopped registering as a scandal and settled into the background of ordinary life, a small ambient dishonesty we all agreed to overlook. You click accept because the alternative is not to use the thing, and the thing is the thing you came to use. The contract is forty pages of adhesion terms drafted by lawyers you will never meet, on behalf of a party whose interests are not yours, and you consent to all of it in the time it takes to find the button. We have made our peace with this. And now we are about to hand the whole ritual to a machine, at a speed and scale that will make the old dishonesty look quaint.

An agent acting on your behalf will encounter these gates constantly. It will sign up for services, accept updated terms, click through consent dialogs, agree to data-processing addenda, and enter into small commercial arrangements — dozens of them in the course of doing what you asked, hundreds over a month, in the ordinary background of getting things done. Each one is a contract. Each one binds you. And the agent will resolve each one the only way it can: by accepting, because acceptance is the path forward and refusal is the dead end. It solves a real friction. It also creates a problem we do not yet have the vocabulary to name.

The contract no one read

Start with what is genuinely being solved, because it is real. The consent theater of modern digital life is a burden no person can discharge honestly. To read every agreement you are asked to accept would consume more hours than the agreements are worth, and everyone involved knows it — the drafters least of all expect you to read. An agent that clicks through on your behalf is, in one sense, merely automating a fiction that was already hollow. The human never read the terms either. Delegating the click to a machine changes the actor but not, at first glance, the substance.

Except the substance does change, in three ways that compound. The first is scale. A person confronts these gates a handful of times a week and is at least dimly aware of doing so. An agent confronts them at machine rate, silently, as a routine step in a larger task, and binds its principal each time without the principal ever knowing a contract was formed. The friction that used to slow us down was also, quietly, a form of notice. Remove it and you remove the one moment where a human might have looked up.

The second is comprehension. When a person clicks accept without reading, they at least possess the capacity to have read — the failure is one of attention, not of kind. The agent has no such capacity in the sense that matters. It can parse the text; it can even summarize it. But it cannot weigh what an obligation will mean to you, cannot know that this particular arbitration clause forecloses a dispute you would have cared about, cannot feel the difference between a term that is boilerplate and a term that is a trap. It processes the agreement; it does not undergo it.

The third is authorship. These are not negotiated contracts. They are adhesion terms, written by the counterparty to favor the counterparty, offered on a take-it-or-leave-it basis. The agent does not push back on a term, because there is no mechanism to push back and no party on the other side prepared to listen. So you end up bound by obligations that were drafted against your interest, accepted by a proxy that cannot judge them, at a volume that guarantees no one on your side will ever review the record. Every structural safeguard that once made consent mean something has been removed at once.

Consent no one gave

This is where the familiar question turns out to be the wrong one. The instinct, when an agent binds its principal, is to ask whether the agent had authority to do so — whether the delegation was valid, the scope was granted, the click was within the mandate. That question matters, and it has answers. But it is not the question that should worry us, because it can be answered yes and leave the real problem entirely untouched.

The real problem is legitimacy, which is a different thing from authority. An agent can be fully authorized to accept terms on your behalf and still produce an agreement that no legitimate act of consent stands behind. Consider the two parties to the transaction on your side. There is you, the principal, who is bound — and who never saw the terms, never engaged with them, has no idea what was agreed. And there is the agent, which did engage with the terms in the mechanical sense of processing them — but which cannot consent in any meaningful way, because consent requires understanding what one is giving up and being able to weigh it, and the agent can do neither. The party who is bound did not agree. The party who agreed cannot mean it. Between them, the act of meaningful consent that is supposed to make an agreement legitimate simply never happens.

An agent can accept the terms in a millisecond; it cannot mean them, and a contract nobody meant is a signature in search of a person to blame.

Authority tells you the agent was permitted to click. It does not tell you that the resulting obligation rests on anything a human would recognize as agreement. This is the gap. We have built systems that can generate authorization at scale — the agent had the mandate, the click was valid — while quietly emptying out the thing authorization was supposed to be a proxy for. A signature used to be shorthand for a person considered this and assented. When the signing is delegated to something that cannot consider and cannot assent, the shorthand still executes, but the thing it stood for is gone. What remains is a binding with no one behind it: authorization without legitimacy.

A record and a boundary

Two things, at minimum, are owed to anyone whose agent will be entering agreements on their behalf. The first is a record — not of the terms alone, but of the act: what was accepted, on whose behalf, under what authority, and at what moment. This is the difference between an agreement that vanished into the background and one that can be brought back into the light. If your agent bound you to something, you should be able to reconstruct that it did, retrieve what it agreed to, and see the authority it was acting under when it did. A binding you cannot even locate afterward is not a binding you consented to; it is one that happened to you. The record is what converts a silent, machine-rate act into something you retain standing to contest.

The second is a boundary. Not every consent is consequential, and it would be absurd to drag a human back into every cookie banner. But some terms bind in ways that matter — that waive rights, commit money past a threshold, license data, or foreclose recourse — and for those, the machine's convenience is not a good enough reason to bypass a human entirely. A consequential term should meet a boundary that requires actual human engagement before it binds: a point at which the agent stops, surfaces what is at stake, and waits for a person who can understand the obligation to accept it or decline. This is not a failure of automation. It is the recognition that some agreements are supposed to cost the assenting party a moment of genuine consideration, and that removing the moment removes the legitimacy along with the friction.

The distinction that matters, in the end, is between convenient automation and legitimate consent. They are easy to conflate, because a smooth agentic flow that binds you to a hundred agreements without bothering you feels like a feature, and mostly it is. But a system that can bind its principal at machine speed, to terms drafted against them, through a proxy that cannot weigh what it accepts, and leave no record and no boundary behind, is not delivering consent. It is manufacturing authorization and calling it agreement. The friction we were so glad to be rid of was doing a quiet job — it was the last place a human could refuse. If we take it away entirely, we should at least know what we gave up when we clicked.

— Dispatches · Summit Cognitive