From Agentic Commerce to Agentic Markets

Agent-to-business commerce today runs one direction. An agent searches, compares, negotiates, and purchases on behalf of a buyer, and open standards now exist for exactly that flow — Stripe and OpenAI’s Agentic Commerce Protocol defines how an agent completes a purchase from a business, with a person still deciding to buy and the agent managing the checkout. On the seller’s side of the same transaction, the price may already move on a signal, but the resource behind it — the SKU, the pallet, the hour of capacity — remains a listing the buying agent reads, not a position another agent can act on. An Agentic Market is a market in which supply-side resources — inventory, production capacity, logistics capacity — are represented by agents that continuously price and reallocate them in response to changing conditions, rather than agents acting only on behalf of buyers. Arco projects this as the next structural wave of agentic commerce, not a claim about a market that already operates this way at scale.

What the demand-side wave already proved

Memo #13: The Machine-Readable Business argued that a business has to expose a Machine-Readable Interface to be discoverable and transactable by agents at all. Stripe’s ACP Completes the Agent-Readiness Stack observed the transactional layer of that stack standardised: an agent can now complete a purchase from a business through a shared protocol, and the businesses that need six months of integration to support it are the ones whose commerce stack was built for a human buyer. Agentic Consumption Pricing — a pricing model in which cost scales with agent call volume and the capability tiers exposed to agents, shifting the unit of consumption from human seat to agent transaction — is how the seller meters that interface. All three describe the demand side of agent-to-business commerce: an agent buying, at a price a human already set, from a catalogue a human already built. None requires the seller to change anything about how the resource itself is tracked or allocated internally. The interface is new. The supply behind it is not.

The obvious objection is that the supply side is already automated, and in two of three senses it is. Dynamic pricing has run for decades — airline yield management, advertising exchanges, energy dispatch — adjusting a number against a signal on a cadence a human reviews. Inventory visibility has run nearly as long: EDI inventory advices, GS1 event standards, and marketplace seller APIs already expose stock positions to trading partners as machine-readable feeds. Neither is what this memo names. A repricing engine moves the price and leaves the resource where it was. A visibility feed lets an outside system read the position and leaves the decision about it with the person who owns it. An Agentic Market requires the third thing: the resource itself — the unit of inventory, the hour of capacity — addressable by an agent that can commit against it and reallocate it. Reallocation, not repricing and not visibility, is the projection.

What addressability means, and what the supply side still lacks

The shape of the interface event the supply side would need is visible in an adjacent domain. Anthropic’s Model Hardware Standard research preview standardised read and write primitives for laboratory instruments and collapsed integration from weeks to hours. It is a driver standard, inside a research preview, and it exposes nothing about finished-goods inventory or warehouse occupancy to any buying agent. Arco reads it as an analogy for the class of interface event that has to happen, not as evidence the event has happened on the supply side — the same reading the Perspective on MHS gave it for hardware. What that class of standard would have to do for inventory and capacity is stricter than what visibility feeds already do: expose a position an external agent can query and commit against, at an integration cost low enough that representing the resource with an agent is cheaper than the inefficiency the human cadence was creating.

The distinction is not cosmetic. A visibility feed answers "what did the resource look like as of the last count." An addressable position answers "what is available right now, what changed since the last query, and can I commit against it." The first supports a human deciding once a week whether to discount a slow-moving batch. The second supports an agent deciding continuously — which only matters if the cost of acting continuously is lower than the cost of the delay, the same arbitrage logic that justifies any agentic deployment, applied here to a market position instead of a workflow.

What a supply-side agent does with it

Not every action a supply-side agent could take carries the same authority, and the gate has to be set before the loop runs. Repricing a slow-moving batch is reversible and cheap to get wrong. Rerouting a shipment mid-transit or committing production capacity to a new order is not. Memo #119: The Threshold Was Never Priced for Irreversibility named the rule: the Physical Intervention Threshold, an Intervention Threshold calibrated by the reversibility of an action rather than by frequency alone — set by the cost of a single wrong action, not how often it occurs. It applies to a market position exactly as it applies to a laser or a liquid handler. A supply-side agent is free to propose, simulate, and reprice continuously, and is gated more tightly the closer an action gets to something that cannot be cheaply undone.

Inside that gate, consider a supply-side agent representing an unnamed manufacturer’s finished-goods inventory. It does not wait for a category manager’s weekly report to notice a batch approaching expiry, a warehouse running under-utilised, or a regional demand shift leaving a surplus in one market and a shortage in another. It reprices and bundles on its own decision cycle, because those actions clear the threshold. It proposes reroutes and capacity commitments to the Steward, or executes them only where their Rollback Cost has already been priced into a tight threshold, because those actions do not. None of this requires new economic theory — dynamic pricing already prices to signal. What changes is the cycle time, from a human-reviewed cadence to the agent’s own decision latency, and what the agent acts on, from a number in a pricing system to the position itself.

The Operator’s Verdict

A market where only the buyer’s side is agentic is half a market. The seller’s price may already move in seconds, but the decision about the resource behind it — what to hold, reroute, or commit — still waits for the human whose review cadence sets its speed, no matter how fast the buyer’s agent gets. Closing that gap does not require a new economic model — dynamic pricing already prices to signal, and inventory feeds already report position. It requires the resource itself to be addressable: a position an external agent can query and commit against, not a number in a pricing system or a line in a feed.

An operator building toward this should treat inventory and capacity positions as candidate agent-addressable interfaces now, on the same footing as the demand-side interfaces already exposed to buying agents, and should score every action those interfaces permit on both axes — how often, and how expensive if wrong — before the first one runs.

Technology changes what is possible. Addressability determines what becomes a market.