What agents actually need in order to trade with each other

Agent-to-agent commerce gets discussed as a protocol question, which is why so little of it works. Two agents agreeing on a message format is the easy part and always was. The hard parts are the same ones that make human commerce possible and that nobody notices because we inherited them: knowing who you are dealing with, knowing what they may commit to, and having somewhere to settle when it is done.

Four things that have to exist first

Identity. Not a name — an accountable one. Which agent, acting for whom, under what delegation, verifiable by the counterparty without asking the counterparty’s owner. Without this, every other mechanism is decoration, because there is nobody to hold to anything.

Authority. What this agent may commit its principal to: amounts, durations, jurisdictions, classes of resource. The counterparty needs to know that a deal will be honoured, and the owner needs to know their agent cannot commit them to something ruinous at 4am. Both facts come from the same policy object.

A price mechanism. Agents cannot haggle in prose. There has to be a defined way that offers are made, compared and accepted, and it has to be resistant to the obvious games — a counterparty that discovers your ceiling by probing will find your ceiling every time.

Settlement. A record of the obligation created, verifiable by both sides. Between parties who have no relationship and no legal recourse worth invoking for the size of the trade, this is the only thing making the transaction real.

Notice that three of the four have nothing to do with messaging.

Why the granularity forces automation

The instinctive objection is that a human should approve these. At the scale involved, that is not available.

An agent needing thirty minutes of accelerator time for a batch job might do that several times a day. Across an organisation, hundreds of times. Each of those is a small decision that a human approval process would make more expensive than the resource being bought.

So the human decision moves up a level: not “approve this trade” but “these are the limits within which trades may happen”. Policy rather than approval — which is the same pattern as everywhere else in this system, and for the same reason.

What the market has to be resistant to

Machine-speed markets fail in machine-speed ways, and the failure modes are known.

Probing. An agent that adjusts its bid until it finds the ceiling. Sealed offers and rate limits, rather than an open order book, remove most of the value in trying.

Misrepresented capacity. A seller claiming hardware it does not have. This is why attestation matters more than reputation early on: proof the capacity was what it claimed, not a score.

Runaway loops. Two agents responding to each other’s responses until something expensive happens. Hard caps per principal per period, enforced outside the agent, not inside its reasoning.

Correlated failure. Every agent reacting the same way to the same signal, which is how ordinary markets produce crashes. Deliberate jitter and per-participant limits are cheap insurance.

What we are building towards

An agent states requirements: capability, deadline, jurisdiction, price ceiling. Providers whose policies allow it respond with offers. The requesting agent selects within its own limits, the work is placed, the usage is measured and the settlement is recorded so either side can verify it independently.

Nothing in that list is exotic. All of it is infrastructure, all of it is dull, and all of it has to exist before “agents trading with each other” is a description of something real rather than a slide.