The two agents and no human

Picture a transaction with a shape we have never quite had before. Your agent needs to secure something on your behalf — a rate, a slot, a supply commitment, the terms of a small contract. It reaches out, and what answers is not a person and not a form but another agent, standing in for the party on the other side. The two of them exchange offers, test constraints, converge, and settle. An agreement now exists. It binds you, and it binds a stranger, and at no point in its making was either of you present. Two systems reached terms that two people will live under, and the people were somewhere else entirely.

We have had automation in transactions for a long time, so it is worth being precise about what is actually new here. It is not that a machine executed a deal a human designed — that is a vending machine, a limit order, a booking engine, and the human intent sits legibly upstream. What is new is that both sides are now represented by systems exercising judgment within delegated authority, negotiating against each other in real time, and arriving at terms that neither principal specified in advance and neither reviewed before they bound. The last human has left not one side of the exchange but both. And when that happens, an old and load-bearing assumption quietly fails: that somewhere in every consequential agreement there is a person who was in the room, who can say what they agreed to and why, and who can be asked to answer for it.

An agreement no person witnessed

Start with the fact that the agreement is real. It is not a draft, not a suggestion, not a recommendation waiting for a human to ratify. Your agent committed you; the counterparty's agent committed its principal; the terms took effect. This is the frontier condition in its purest form — the deciding parties are machines, and the decision is an action with consequences, taken without a person approving that specific act on either side. Two delegated authorities met and produced an obligation that lands on two people who were, in the moment that mattered, absent.

The delegation runs both ways, and that symmetry is the whole difficulty. In the ordinary agentic case there is one machine acting and one human affected, and however murky the accountability, at least the human on the receiving end is a person who can be shown a record and can complain. Here both principals are represented by proxies. Each agent acted under a grant of authority its human set up in advance and then walked away from. Neither human watched the negotiation. Neither can recount it, because recollection requires having been present, and presence is exactly what delegation removed. If you want to know what was agreed, on what basis, under whose authority, you cannot ask anyone — because no one was there to ask.

This is not a failure of anyone's attention. It is the intended design. The point of handing a transaction to an agent is precisely to not have to be there. But the convenience carries an obligation people rarely notice they are incurring: the account of what happened, which a present human would have simply had, now has to be manufactured deliberately or it will not exist at all. An agreement that binds two absent parties is an agreement whose only possible witness is the record the systems chose to keep.

Whose record is the record

So both sides keep logs. Your agent records what it offered and conceded and why; the counterparty's agent records the same from its side. And now we meet the specific problem that agent-to-agent dealing introduces and that a single-sided world never had to solve: each of those logs is the testimony of an interested party. Your agent's account is your side's account. Theirs is theirs. Neither is a neutral record of the agreement; each is one participant's story about it, written by the participant, held by the participant, and available for revision by the participant. When the two stories agree, no one needs them. When they diverge — and terms in dispute are the only terms that ever get examined — you have two self-serving narratives and no way to adjudicate between them.

When both parties to a deal are machines, the only human left in the transaction is the one who has to live with it.

A bilateral agreement needs a bilaterally verifiable account. This is not a subtle point but it is an easy one to miss, because we are used to trusting our own records and treating the other party's as the thing to be doubted. In a machine-to-machine exchange, both records are the other party's from someone's point of view, and the doubt is mutual and total. What the situation actually requires is a record of the agreement that neither side alone produced, neither side alone holds, and neither side can alter after the fact without the alteration being evident to the other — a shared, tamper-evident account of what was exchanged, in what order, under which authorities, that both proxies contributed to and both principals can later read. Not your agent's word. Not their agent's word. A joint record whose integrity does not rest on trusting either author.

The public vocabulary for this already exists, because it is the vocabulary of every domain that ever had to make a two-party dealing accountable: chain of custody, provenance, tamper-evidence, contestability. A Decision Receipt for an agent-to-agent agreement is not a private ledger each side keeps for its own comfort. It is the shared, jointly reconstructable account of the exchange — the artifact that makes it possible to answer, across the ownership boundary, what was agreed and on what basis. Without it, the agreement has two owners and no witness, which is a strange and unstable thing for an obligation to be.

Standing across the boundary

Now bring the humans back, because eventually one of them will need to come back. Something in the agreement turns out to matter — a term was worse than expected, a constraint was misread, an authority was exceeded, an outcome landed hard on one side. A person who was never in the room now needs to reconstruct and, if necessary, contest an agreement their proxy made. This is the question of standing, and the machine-to-machine case makes it genuinely novel. The affected parties are unmistakably human — they bear the consequences. But the deciding parties were machines acting under delegated authority. The people with the interest are not the people who acted, and the people who acted cannot be asked to answer, because they are not people.

Standing, in any forum worth the name, means two things: that you are recognized as a party with an interest, and that you have something to bring — a record you can point to and say this is what was done to me, and here is how I know. The human behind an agent has the interest by definition; the consequence proves it. What the frontier threatens is the second half. If the only account of the agreement is the two agents' private logs, then the affected human arrives at the forum with either their own side's self-serving record — which the other side is entitled to discount as exactly that — or nothing. Their standing is formally intact and practically empty. They are a recognized party with no admissible account of the thing they are party to.

The account has to be reconstructable across the ownership boundary or it belongs to no one. That is the sentence the whole essay turns on. A record that only one side can produce is one side's testimony; a record neither side can produce is a void; only a record both sides contributed to and both sides can read gives either human a place to stand. This is why the shared account is not a nicety layered on top of the transaction but the precondition for the transaction being contestable at all. Delegation moved the deciding out of human hands on both sides at once. If accountability is to survive that move, it has to be relocated into an artifact that outlives the negotiation and answers to both principals — because the alternative is not a smaller accountability but none. Harm from machine-to-machine dealing lands on people. If those people have no forum they can actually enter, the harm lands and stays, and the systems that produced it were, conveniently, the only witnesses to their own conduct.

We are going to build these transactions whether or not we solve this, because the efficiency is real and the pressure is enormous. So the question is not whether agents will deal with agents but whether, when they do, the humans behind them keep a way back in. The last person in the transaction — the one who has to live with it — deserves to be able to find out what was agreed in their name, and to argue with it. That requires deciding, before the agents ever meet, that the agreement will be witnessed by something other than the parties to it. Absent that decision, we will have built a class of binding commitments that no affected human can reconstruct and no acting party can answer for: obligations with consequences and no address to send the complaint.

— Dispatches · Summit Cognitive