The deal two agents struck

Two agents meet — not in a room, because they have no rooms, but across an interface — and in the time it takes a human to reread a sentence they settle a price, a delivery date, a set of terms, and part ways. Behind each of them stands a principal: a person, or a firm, who will be bound by whatever the two of them agreed. The buyer's agent conceded a little on timing to hold the price; the seller's agent accepted a penalty clause it read as narrow and the other read as broad. The deal is done. And the two humans who now owe each other something were not present for a single word of it.

This is not a distant scenario. It is the plain consequence of letting agents transact on our behalf, which is among the first things anyone wants agents to do. Booking, procuring, scheduling, contracting — these are delegations we make gladly, because they are tedious and an agent is tireless. But delegation of a negotiation is not like delegation of a search. When an agent negotiates, it does not merely act; it commits its principal to the acts of another party. The output is not information for me to use. It is an obligation I now carry, created by a conversation I did not witness.

The machine-brokered agreement is a new accountability object, and it arrives with all the classic problems of contract wearing unfamiliar clothes. What was agreed? By whom, with what authority to agree it? And when the two sides later disagree — as parties to consequential deals reliably do — against what account is the disagreement resolved? These are old questions. What is new is that the only witnesses to the answer are two machines, each keeping its own books.

The agreement neither principal saw

Start with the strangeness of being bound by terms you never read. In ordinary commerce, the person who signs is, at least in principle, the person who saw the document. The signature is a claim: I have read this and I assent. Agentic transactions sever that link. My agent assents on my behalf, at machine speed, across possibly hundreds of small concessions, and the first I hear of the finished shape of the deal may be a one-line summary after the fact — if that. I am bound not by what I read but by what my delegate did in my name.

Two further gaps open underneath. The first is authority. My agent had some mandate from me — a budget, a set of acceptable terms, a goal — but the boundary of that mandate is rarely crisp, and the counterparty cannot see it at all. Did my agent have standing to agree to that penalty clause, or did it exceed what I would have allowed? If the deal binds me, that question is not academic; it is the whole question. An agreement reached by an agent acting beyond its authority is an agreement I have grounds to disown — but only if there is a record of what the authority was.

The second gap is memory. Each agent walks away with its own account of what happened: its own log, its own summary, its own interpretation of the ambiguous clause. There is no reason these two accounts agree, and every reason they will diverge precisely where the money is. Each principal, asking its own agent what was decided, receives a confident and internally consistent answer — and the two answers do not match. Neither side is lying. They are each reading their own reconstruction of a conversation that was never jointly recorded.

The old contract problem at machine speed

None of the three difficulties — unseen terms, uncertain authority, divergent memory — is genuinely new. They are the permanent problems of agreement between parties who cannot fully see into each other's intentions, which is to say, the problems law has worked on for as long as there has been trade. What is new is the form and the volume. The questions arrive at machine speed, in machine quantity, generated by delegates that do not tire and do not, on their own, produce anything a third party could later adjudicate.

Consider what a traditional contract is, stripped to its function. It is a single, shared artifact that both parties helped shape and both parties hold — the same words, one document, mutually acknowledged. Its whole power comes from being one record rather than two. When a dispute arises, the parties do not compare their private recollections; they return to the shared text. The contract is the neutral thing that outranks both memories.

Agent-to-agent transactions, left to themselves, produce the opposite: two private logs and no shared text. Each side has a full and detailed account, and the very completeness of each account makes the divergence worse, because each party can point to something concrete and say here, this is what was agreed. A disagreement between two confident, incompatible, evidence-backed reconstructions is not a disagreement that resolves itself. It is a dispute with no floor — no authoritative account to which both sides are already committed. And it multiplies at exactly the rate that agentic commerce scales.

Two agents can shake hands in a microsecond; the humans they bound will spend months arguing about what the handshake meant, unless something neutral wrote it down.

A shared record of the deal

The remedy is not to slow the agents down or to keep a human reading every clause — that surrenders the point of delegating at all. The remedy is to give the machine-brokered agreement the one thing the traditional contract had and the two private logs do not: a shared, neutral, tamper-evident record of what was actually agreed, held in common by both sides rather than kept privately by each.

This is the tally stick brought forward. For centuries a debt was recorded by notching a single stick and splitting it lengthwise, each party keeping a half; the agreement could be verified because the two halves had to match, and neither party could alter its own half without the mismatch giving the alteration away. The genius was not the wood. It was that the record was structurally two-party and matched — a single fact of agreement, mutually held, that no one side could quietly rewrite. A machine-brokered deal needs the same property in digital form: one account of the terms, produced at the moment of agreement, that both agents attest to and neither can later revise in private. When the principals disagree, they do not litigate their agents' competing memories. They read the matched record.

And the record must carry more than the terms. It must carry, for each side, the authority under which its agent agreed — the mandate it was operating within at the instant it committed its principal. This is the older idea that an agent's authority is not a private setting but part of the account of the act: what it was permitted to bind, recorded alongside what it bound. With that, the disownment question has an answer. If my agent stayed within its recorded mandate, the deal holds and I am bound whether or not I read it. If it exceeded that mandate, the record shows the excess, and the agreement is contestable on grounds both sides can see. Either way, the dispute is resolved against an authoritative account rather than against whichever party argues its own reconstruction more forcefully.

That is the whole of it. Agents transacting on our behalf is not a problem to be prevented; it is a capability worth having. But a capability that creates obligations must also create the record that makes those obligations answerable. An agreement that binds two absent humans, reached by two delegates keeping two private books, is a dispute waiting to happen with no floor beneath it. Give the deal a shared, tamper-evident account of what was agreed and by what authority, and the same transaction becomes something a third party — a court, an arbitrator, the parties themselves — can actually resolve. The handshake was always going to be too fast to watch. The least we can do is make sure something neutral wrote it down.

— Dispatches · Summit Cognitive