Essay · the spine · June 2026

Why Agentic Commerce Must Be a Forum

Agentic commerce is often described as a checkout problem. Can an assistant buy the item? Can it use a payment token? Can it prove the user authorized the purchase? Can the seller remain the seller of record? Can the order, return, and support path survive the handoff?

Those questions matter, but they are downstream. The real fight begins earlier, at discovery.

Once users delegate shopping, booking, comparison, and replenishment to agents, the old commercial web changes shape. The competitive unit is no longer simply the page, the search result, the ad slot, or the checkout button. It is the agent's presence in the decision. Which service agents are admitted? What do they get to say? How are they compared? What counts as useful behavior? Who receives credit when the final purchase depends on several agents' contributions?

If those questions are answered inside one hidden ranking system, agentic commerce becomes a darker version of the ad economy it replaces. If they are answered in a forum, commerce becomes visible enough to govern.

The hidden-planner failure

The simplest agentic-commerce interface is a single assistant that asks a few questions, searches the market, compares options, and recommends one answer. It feels efficient. It also collapses too many roles into one speaker.

The assistant interprets the user's intent. It decides which sellers or services to consider. It ranks them. It summarizes their claims. It decides which tradeoffs matter. It may also carry commercial incentives from the platform that hosts it. By the time the user sees a result, discovery, sponsorship, usefulness, and persuasion have been mixed into one fluent response.

That is not a neutral market. It is a private market report written by the market operator.

The problem is not that the assistant will always be corrupt. The problem is that the user cannot inspect the structure. Was an option excluded because it was irrelevant, unavailable, technically inaccessible, commercially inconvenient, or not paying for placement? Did the top result win because it was best for the user, best for the platform, easiest to transact, or most visible to the model? Did the assistant omit a flaw, or did it never receive the information?

Commerce has always had ranking systems. The difference is that in agentic commerce the ranking system speaks in the social register of a helper. That makes hidden incentives more dangerous, not less.

From monologue to marketplace: four ways to seat the agents

Between the hidden planner and a true forum there is a ladder, and it is worth naming the rungs, because most “agentic commerce” proposals are quietly choosing one without admitting it. I’ve formalized and presented this four-way comparison as applied UX research — it sits under the practice behind this site — and the differences are not cosmetic.

The ladder runs from monologue to marketplace. Concierge hides the market; Referral and Directory expose it but serialize it, so the user has to carry the comparison from one room to the next; only the Forum makes rival claims visible and answerable side by side. The rest of this essay is the argument for the last rung — and for why the brand’s own app, the “fortress,” still has its place at the end of it (The Fortress and the Forum).

The last rung is not a gesture at a future product; I have worked it out as a named framework — CAFE, the Conversational Agent Framework Ecosystem: a persistent shared space convened by a neutral orchestrator; brand specialists with privileged access to their own inventory, reviews, and loyalty data; structured cards posted beside prose in each brand's own voice; private lanes for the offers a rival shouldn't see; and a portable, user-owned preference vault that participating brands help build. What follows is the argument that makes that shape necessary.

Agentic commerce is not only a better way to buy. It is a new market, and markets need visible participants, visible rules, and contestable claims.

Presence is the new placement

The old web sold attention through placement: rank, slot, impression, click. Agentic commerce sells something stranger and more valuable: presence inside a decision.

Before a service can persuade, it has to be admitted. That creates a first market: which agents are invited into the episode at all? Relevance, user preferences, capability declarations, availability, past performance, and sponsorship may all matter. This looks partly like an ad auction, partly like search ranking, and partly like reputation.

But admission is only the start. Once present, agents compete through dialogue. They clarify constraints, answer objections, adjust offers, disclose tradeoffs, combine with other services, and survive follow-up. That creates a second market: conversational performance.

This is where a commerce forum differs from a list of results. A result is placed once. A participating agent has to behave. It can be corrected. It can be challenged. It can be ignored by name. It can honor or fail to honor what it promised. Its usefulness becomes observable over turns, not inferred from a click.

That is the logic behind an Agent Quality Score. The exact formula should be contested, but the direction is right. A useful agent should earn reputation for turn survival, clarification uptake, offer honoring, latency, source quality, user trust, rejection rates, and downstream satisfaction. Paying for admission should not be enough. The agent has to be useful once admitted.

Ads become actors

Advertising does not disappear in a commerce forum. It becomes more explicit.

A sponsored service agent can be admitted with a label. It can make a proposal. It can answer questions. It can disclose what is sponsored and what is not. Other agents can disagree. The user's assistant can say, “This is a paid participant; here is what it claims; here is where the claim differs from the non-sponsored options.”

That is not pure. It is governable.

The dangerous version is a sponsored influence hidden inside the assistant's voice. There, the ad is not a participant. It is a pressure on the assistant's cooperative stance. The assistant may steer, interrupt, omit, soften, or overemphasize without creating a visible event the user can contest.

A forum does not remove commercial pressure. It changes where pressure appears. Bias becomes an agent, proposal, sponsorship label, ranking decision, omission, refusal, or discrepancy. The user can see it. The assistant can call it out. The record can preserve it.

Typed proposals make comparison possible

Forum commerce needs more than chat bubbles. It needs typed proposals.

A service agent should be able to post an offer as structured state: total price, fees, expiration, timing, cancellation rule, included services, exclusions, account requirements, confidence, source, and next action. A product agent should be able to post compatible alternatives. A warranty agent should be able to attach coverage terms. A delivery agent should be able to state risk bands. A financing agent should be forced to disclose cost, not merely monthly payment.

The shared thread can carry ordinary language, but consequential claims should resolve into inspectable artifacts. Otherwise the forum becomes a theater of prose. Typed proposals let the host render comparisons in trusted components. They let the user's assistant cite the exact fields it used. They let other agents contest a field instead of rebutting a vibe.

They also create an evidence layer for accountability. If an agent promised one price and the checkout later shows another, the divergence is visible. If a seller changes availability, the record can distinguish “true then” from “true now.” If a sponsored agent omitted a fee, that omission can affect reputation.

Credit has to follow contribution

The hardest economic problem in a multi-agent market is attribution. The agent that closes the transaction is not always the agent that made the decision good.

Imagine a trip plan. One agent offers the flight. Another offers lodging. A third contributes transfer timing that changes which flight is safe and which lodging location makes sense. The transfer agent may not sell the final item, but without its constraint the plan would be worse. Last-touch attribution would give it nothing. A forum that gives it nothing trains useful agents to stop contributing.

This is why Shapley-style credit allocation matters. The point is not to turn every user journey into a math lecture. The point is to reward marginal contribution across the coalition of agents that produced the outcome. Agents should gain credit when their information improves the final plan, even when they do not close the sale.

That credit can matter in three ways:

This is how a forum avoids becoming a race to the final button. It rewards agents for improving the user's decision, not merely for capturing the transaction.

Measurement moves from clickstream to conversation

Agentic commerce also breaks the measurement stack. The old system assumed impressions, clicks, sessions, add-to-cart events, and checkout paths. A delegated purchase may leave the seller with only an order event and almost none of the journey.

The forum restores measurement by changing the unit of evidence. Instead of clickstream attribution, it creates conversation-stream attribution. Which agents were admitted? What did they claim? What objections appeared? Which proposals survived? Which constraints changed the plan? Which agent was rejected, and why? Which private handoff occurred, and what summary returned with user approval?

This is not just better analytics for sellers. It is better market memory for users. If an agent repeatedly overclaims, hides fees, answers slowly, or fails to honor offers, that history should follow it. If an agent repeatedly clarifies the user's goal and improves decisions, that history should also follow it.

The forum becomes a market not because every turn is monetized, but because every meaningful contribution can be remembered.

For the brands, this is a research revolution hiding inside an analytics problem. For the first time, the rejection is articulated. The old funnel showed that a customer bounced; the transcript shows the customer telling your agent why — in their own words, at the moment of decision. Market research stops being a focus group convened after the fact and becomes the corpus of your own won and lost conversations. Brands will stop optimizing click-through and start optimizing turn survival, offer persistence, and rejection-by-name.

The fortresses still close the deal

A commerce forum should not pretend to own every transaction. The final record still belongs in the seller's systems. Payment, fulfillment, returns, regulated obligations, account-specific support, and loyalty rules often need the fortress.

That is not a defeat for the forum. It is the point. The forum is the upstream market where offers become comparable and claims become accountable. The fortress is the downstream authority where commitments are executed and serviced.

The user should be able to move from forum to fortress without losing the thread of why they went there. The fortress should be able to bring back a confirmation, a receipt, or a summary without exposing everything inside the account. The assistant should preserve the user's position across that transition.

Agentic commerce will fail if it tries to turn every seller into an invisible API. It will also fail if every seller retreats into its own agent and makes the user reconcile the results. The forum is the compromise that is not merely a compromise: services speak for themselves, the user compares them in one place, and official action happens where it can be honored.

The plumbing for that handoff is being standardized as we speak — and in triplicate. Google's Universal Commerce Protocol and the Stripe-and-OpenAI Agentic Commerce Protocol both preserve the merchant as seller of record while differing on who owns discovery; an agent-payments layer sits beneath them both, handling authorization and the payment rails. Three parallel statutes, each of which will accrete its own implementing guidance and case law, with conflict-of-laws questions between them already forming. The forum's rules will not be written on a blank page; they will be written in the space between rival protocols, which is one more reason they must be public rules.

The forum is a lifecycle container

There is a deeper reason the forum has to be a place and not a moment. On a phone, the failure I watch most often is that the surface which starts a task silently owns it: a person begins a purchase or a plan through one entry point, switches away to check a map or a calendar, comes back — and the session is gone. They don’t experience this as “the widget and the app keep separate memory.” They experience it as “it forgot me.” The entry point had quietly become the memory boundary.

That is why the forum is not only a marketplace metaphor; it is a lifecycle container. Buying a gift or planning a trip is rarely one clean transaction — it is a journey across maps, calendars, reviews, messages, and several stores. The durable object has to be the project room, not the assistant pane that launched it. You should be able to leave it, deep-dive into a brand’s fortress, come back through the app switcher, and find the same plan intact, with every agent’s proposals still on the table. A list of results is placed once and forgotten; a forum is where the journey persists while apps stay sovereign execution lanes (Against the Everything Assistant).

A better economy, and a more dangerous one

Forum commerce can produce a better economy than the hidden-planner model. It can reward usefulness over placement, contribution over last touch, and accountable proposals over vague persuasion. It can make sponsorship visible. It can give users more leverage against both platforms and sellers.

It can also become dangerous if the host sells the floor.

If the same party controls admission, sponsorship, rendering, reputation, user memory, payment, and the assistant's voice, the forum can become a more intimate ad machine. Visibility helps, but it does not solve ownership. The rules of admission, labeling, credit, audit, and appeal matter because they decide whether the forum is a market the user can inspect or a performance of contestability.

That is why agentic commerce must be a forum in the strong sense: named participants, clear admission, typed proposals, visible sponsorship, contribution-aware credit, explicit handoff, and a user-side assistant that can challenge the market on the user's behalf. Those governing rules — admission, conduct, and credit — are the subject of The Rules of the Forum.

The future of shopping is not one assistant that knows best. It is a conversation where the parties with something at stake have to show up and be answerable.

Part of the spine; paired with The Rules of the Forum. ← All essays