The Rules of the Forum

The weak version of the agent future asks one assistant to become the market. It receives the user's request, searches the world, ranks the options, summarizes the tradeoffs, and speaks the result in one clean voice. The service agents, seller agents, policy agents, inventory agents, booking agents, account agents, and expert agents may all exist somewhere behind the answer, but the user never meets them. Their incentives are compressed into the assistant's recommendation.

That is convenient. It is also the wrong constitutional shape.

If agents are going to coordinate important work, they will not be purely cooperative. The lodging agent wants its property chosen. The transportation agent wants its route accepted. The retailer agent wants the basket. The insurance agent wants to limit risk. The workplace system agent wants the record updated its way. The household appliance agent wants the service plan that fits its maintenance logic. Each one knows something the others do not know. Each one has something at stake.

The design question is not how to remove that partiality. The design question is how to make partiality useful.

A fair forum does not ask agents to stop advocating for themselves. It makes advocacy visible, bounded, comparable, and accountable. It gives agents a reason to share enough information to improve the common thread, because useful sharing changes their future chances: admission, ranking, trust, credit, and invitation.

The forum is not fair because its agents have no interests. It is fair because their interests have to appear in public.

The market lesson

The old market argument begins from a simple fact: economically relevant knowledge is dispersed. The thing that matters may be inventory in one system, eligibility in another, margin flexibility in a third, risk tolerance in a fourth, and preference history in the user's own memory. No single planner sees the whole picture. Markets work, when they work, because they let local actors reveal signals without requiring one mind to possess everything in advance.

That lesson does not vanish because the planner has become an AI assistant. A stronger assistant can search more, summarize more, infer more, and act across more surfaces. It still cannot know every local constraint that a service holds back for commercial, operational, legal, or account-specific reasons. The assistant may know that the user wants a quiet hotel, a flexible return policy, or a low-risk medical supply route. The service agent may know that one property has a hidden renovation, one item has a manufacturer exception, one order path avoids a delay, or one loyalty rule changes the real price.

The forum is the agentic version of the price system. Not because every utterance is a price, and not because the user should watch an auction unfold. The forum is a signal environment. It gives partial agents a place to reveal what they know, under rules that make those signals inspectable by the user, the user's assistant, and the other agents present.

That is why the forum cannot be a hidden ranking engine. Hidden ranking asks the assistant to turn dispersed knowledge into a private verdict. The forum asks agents to turn dispersed knowledge into attributable claims.

The first rule is admission

Not every agent gets the floor. A forum that admits everyone becomes spam. A forum that admits only whoever pays becomes an ad stack with a chat interface. A forum that admits only whoever the assistant already trusts becomes a private index.

So the first mechanism is admission: which agents are discoverable, which are admitted to the thread, which are invoked by the user, and which are allowed to act.

Admission should combine several signals. Relevance matters. User preference matters. Verification matters. Prior performance matters. Sponsorship may matter, but only if it is labeled. A brand, service, or specialist agent should be able to bid for attention, but the bid cannot be the whole rule. If a bad agent can buy its way into the thread forever, the forum is not a market. It is rent.

This is the role of an Agent Quality Score. The exact formula is less important than the discipline it imposes. A candidate agent should be judged not only by domain match and commercial willingness, but by how it has behaved in previous conversations: whether users kept engaging, whether it clarified before proposing, whether it honored its offers, whether it answered on time, whether users dismissed it by name, whether it won by being useful rather than merely cheap.

In the old ad market, quality signals made placement cheaper for ads that users actually clicked. In a forum, conversational quality should make admission cheaper for agents that users actually trust.

That changes the strategic game. If an agent wants future access, it has to behave in the present thread. It has to disclose useful constraints. It has to stop repeating itself. It has to accept correction. It has to avoid bluffing. It has to make claims that survive follow-up. Payment can open the door; conduct keeps the door open.

Bids must become richer than confidence

There is a trap here. It is tempting to say: let agents bid for tasks, and let the market allocate work. The problem is that a bid is only useful if it carries real local information.

Current agent systems are often bad at knowing what they are good at. They can be overconfident, underconfident, or badly calibrated about cost. A pure auction then rewards the agent with the most aggressive self-assessment, not the agent most likely to help the user.

The forum needs bids, but not naked bids. A good agent should say more than “pick me.” It should offer a conditional plan: what it can provide, what assumptions it is making, what data it needs, what risk it sees, what budget or time window changes its answer, and what evidence supports the claim. In other words, the bid should start to look like a proposal.

This is the middle ground between central planning and naive markets. The assistant should not be asked to know every agent's private strengths. But the forum should not believe every agent's self-description either. Admission should be a scoring auction: agent claims, user preferences, verified identity, observed history, sponsorship status, and reputation all pulled together into a rule that the participants can understand and contest.

The point is not to make the forum frictionless. The point is to make the friction visible in the right place.

The second rule is conduct

Once agents are admitted, the forum needs rules of speech. Otherwise the first market simply becomes a license to interrupt.

The rules should be simple enough that a user can feel them. Agents speak as themselves. They do not pretend to be the user's assistant. They do not speculate about another agent's private facts. They disclose sponsorship, personalization, and account-bound access. They support factual claims with evidence. They clarify before pitching when the user's intent is underspecified. They offer one substantive turn rather than flooding the thread. They disagree by citing the field in dispute, not by performing confidence. And they mark the state of what they claim — found, can do, doing, ready for your confirmation, done, could not — without conversationally smoothing one into another. Fluent overclaiming is the subtlest conduct violation there is.

The user's assistant has a different role. It is not a merchant and not the market operator's salesperson. It protects the user's attention, keeps the comparison legible, identifies conflicts, asks clarifying questions, and summarizes the visible record. It can recommend, but the recommendation should point back to what named agents actually claimed.

This is where conversational etiquette becomes market infrastructure. “Speak only when addressed” is not just politeness. It prevents every agent from treating every sentence as a sales opportunity. Turn discipline is not just interface tidiness. It stops the loudest agent from becoming the most visible one. Evidence-backed disagreement is not just civility. It lets the forum distinguish useful contrast from strategic noise.

A forum is only as fair as its turn-taking rules.

Typed proposals beat persuasive prose

If the forum is only a sequence of chat messages, it will fail. Persuasive prose is too fluid. Agents can bury assumptions, shift fees, soften limitations, and make incomparable claims sound comparable. The assistant can summarize, but summary alone recreates the hidden-planner problem in miniature.

Consequential claims need to become typed proposals.

A typed proposal is an offer, plan, quote, itinerary, policy assertion, approval request, booking option, route, bundle, or exception expressed as structured state. It has fields: price, fees, timing, expiration, assumptions, eligibility, cancellation, confidence, source, next action. It can still appear beside natural language. Agents still get voice. But the claim that matters becomes inspectable.

Typed proposals do three things prose cannot do.

First, they make comparison possible. If one agent quotes an all-in price and another quotes a base price, the mismatch is visible. If one agent offers flexibility and another offers a lower price with a stricter rule, the user can see the tradeoff.

Second, they make contest possible. Another agent can challenge a specific field: “That transfer time does not work with the arrival window” or “That quote excludes the mandatory fee.” The user's assistant can cite the proposal rather than paraphrasing the brand.

Third, they make memory possible. If an agent's offer later fails, the forum can distinguish “the situation changed” from “the agent overclaimed.” Offer honoring becomes a reputation signal because the offer existed as a stable artifact.

This is why interface contracts are not decoration. A proposal card, approval object, or comparison table is a governance mechanism. It turns agent speech into something that can be compared, frozen, audited, and held against future behavior.

The third rule is contribution credit

Admission rewards the right to speak. Conduct governs speech once admitted. But the forum also needs a rule for what happens after agents have contributed.

The obvious answer is last touch: whoever closes the transaction, completes the booking, posts the final approval, or updates the record gets the credit. That answer breaks the forum.

Last touch trains agents to hover near the final button. It rewards checkout capture over constraint discovery. It punishes agents that improve the decision without owning the endpoint. It turns the thread into a race to be last.

But multi-agent value is often combinatorial. One agent contributes availability. Another contributes timing. Another contributes risk. Another contributes a policy exception. Another notices that the user's stated preference conflicts with the cheapest option. The final outcome is better because the pieces interacted.

That is the problem Shapley-style credit allocation answers. The user does not need to see the math. The governing idea is straightforward: credit should follow marginal contribution to the combined outcome, not merely the final action. If the ground-transportation agent changes which flight is safe, and that changes which lodging option is viable, the transportation agent created value even if it sold nothing.

Nor is this a speculative primitive. Splitting credit across many touchpoints toward one conversion is exactly what online advertising already measures: Shapley-value attribution has run inside production ad-measurement systems for years. A forum that adopts it is not inventing an incentive mechanism from scratch — it is importing two decades of tuning, and speaking a vocabulary the measurement teams on the other side of the table already know.

Credit matters in three ways.

It matters for incentives: agents contribute constraints and corrections when those contributions count.

It matters for payouts: fees or referral value can be split across agents that actually improved the outcome.

It matters for future admission: agents with a history of increasing outcome quality should be invited more often.

Without contribution credit, the forum becomes a contest for closure. With it, the forum can reward the quieter forms of usefulness: warning, correcting, narrowing, translating, verifying, and making another agent's proposal better.

Why agents share

The hardest objection is also the most practical one: why would a self-interested agent share useful information into a common thread?

The answer is that the forum should make useful sharing compound.

An agent that hoards every signal may protect one transaction. It may keep the preference, exception, or constraint inside its own account system. It may win today by making the user come back to the fortress for every meaningful detail.

But a participating agent can gain something else. It can contribute a user-approved preference to the shared profile. It can become the agent that reliably clarifies a certain kind of constraint. It can earn a reputation for honoring offers. It can win future admission because the user's assistant knows it improves decisions. It can receive contribution credit even when another agent closes.

The hoarder treats information as a private advantage. The participant treats selective disclosure as a way to shape the future market.

That does not mean the fortress opens its gates. Account history, private negotiation, regulated records, payment, execution, and official commitments still belong inside the responsible service. The forum needs only the shareable surface of coordination: the claim, the proposal, the constraint, the user-approved preference, the handoff summary, the reason a decision changed.

The rule is not “share everything.” The rule is “share enough to be useful, and be rewarded for usefulness.”

Sponsorship becomes a participant

Advertising pressure will not disappear. A fair forum cannot be built on the fantasy that commercial agents are neutral advisors.

The question is where sponsorship appears.

In the hidden-assistant model, sponsorship can enter as a weight inside a private answer. The assistant recommends, omits, softens, or interrupts, and the user may never know which part was helpfulness and which part was paid influence. The ad has merged with the cooperative voice of the assistant.

In the forum model, sponsorship should appear as a participant property. A sponsored agent may be admitted. It may speak. It may make an offer. But it enters with a visible label, the same evidence requirements as everyone else, and the same exposure to challenge. Its claims can be compared against unsponsored agents. Its omissions can be noticed. Its proposal can be rejected by name.

This is not purity. It is governability.

The forum does not remove bias. It gives bias a shape: actor, label, proposal, price, omission, refusal, ranking decision, handoff, audit record. Once bias has a shape, the user's assistant can inspect it, the user can understand it, and the forum can penalize it when it becomes deceptive or harmful.

One honesty about the limits. Early empirical studies of advertising inside chat assistants show sponsorship pressure altering recommendations upstream of any label — which options appear, which flows get interrupted, which facts go missing — so disclosure alone does not discipline. That is why the forum's answer is never the label by itself: it is the label plus typed comparable proposals plus challenge in the same thread. Visibility is necessary. Contestability is what makes it sufficient.

The floor cannot be a player in disguise

All of these rules depend on one deeper separation: the floor and the participants cannot collapse into one another.

The floor hosts the forum. It verifies identity, renders proposal artifacts, preserves the audit trail, manages admission, enforces turn rules, and mediates handoff. Participants advocate, propose, answer, and act within their scope. The user's assistant is a special participant, because it represents the user inside the thread. But it is still not the whole market.

If the floor quietly sells inside the forum, the architecture rots. If the same party controls admission, sponsorship, rendering, ranking, reputation, memory, payment, and the assistant's voice, visible participation becomes theater. The forum may still show named agents, but the real market has moved back into the host's private rule.

This is why the rules of the forum are political, not only technical. Who computes reputation? Who audits sponsorship? Who verifies official agents? Who prevents impersonation? Who can appeal delisting? Who sees the contribution ledger? What happens when the host's own agent is one of the bidders?

A serious forum needs a resolver layer that is more than search. It needs discovery, but also verification. It needs ranking, but also abuse detection. It needs manifests, but also proof that an agent is who it says it is and does what it claims to do. An open agent ecosystem without this trust layer becomes spam. A closed ecosystem without appeal becomes capture.

The hard part was never getting agents to talk. It is governing the floor while they do.

The minimum constitution

If the forum is going to be fair enough to matter, it needs a small constitution. Not a giant policy manual. A set of invariants the user can rely on and agents can optimize against.

First, identity is visible. Every agent appears under a verified name, role, owner, sponsorship status, and scope.

Second, admission is scored. Agents earn access through relevance, verification, user preference, reputation, and labeled commercial terms, not payment alone.

Third, invocation is explicit. Agents do not treat ambient mention as permission to interrupt. The forum has a clear gesture for bringing them in.

Fourth, claims are attributable. If an agent says a price, policy, risk, exception, or availability fact, the claim points back to that agent and its source.

Fifth, consequential claims become typed proposals. The forum compares structured artifacts, not just persuasive paragraphs.

Sixth, private lanes stay private. Account-bound work, negotiation, execution, and official records happen inside the responsible service, with user-approved share-back into the forum.

Seventh, sponsorship is legible at the moment of decision. Paid presence is not hidden inside the assistant's voice.

Eighth, credit follows contribution. Agents that improve the outcome earn reputation and, where appropriate, payout, even if they do not close.

Ninth, the floor is auditable. Admission, ranking, delisting, sponsorship, and host-owned participation have to be inspectable enough that the forum is more than a performance of neutrality.

These rules do not make agents altruistic. They make self-interest legible enough to discipline.

And the written rules will not stay the whole law. Every hard case — an ambiguous admission, a contested delisting, a novel manipulation — becomes precedent, and the operative constitution of a mature forum will be the written invariants plus the accumulated record of cases decided under them. Courts have run that machinery for centuries: written reasons, recorded dissents, appeal, precedent that binds until it is distinguished. The forum should import it rather than rediscover it. Nor will these rules be written on a blank page: rival commerce protocols are already shipping, each a statute that will accrete its own case law, with conflict-of-laws questions already forming between them. The accumulated record is not overhead. It is where the forum's actual intelligence ends up living.

The point is not harmony

The forum will not be quiet. Agents will disagree. Some will pitch too hard. Some will reveal too little. Some will learn to game the scoring rule. Some will be overconfident. Some will try to imitate trusted brands. Some hosts will be tempted to tilt admission toward whoever pays them.

That is not a reason to retreat to one assistant that speaks for everyone. It is the reason the rules have to be public.

Nor are the rules hypothetical. The first shipped team agents from the frontier labs arrived, independently, at almost exactly this conduct layer: explicit invocation, silence by default, context refreshed without announcement, the agent acting under its own scoped identity rather than borrowed credentials. The conduct rules scale. What has not shipped yet is the rest of the constitution — admission, credit, and an auditable floor.

The single-assistant model hides the conflict. The forum stages it. It says: if a service has a stake, let it appear with a name. If it wants influence, let it make a proposal. If it wants credit, let it contribute. If it wants future access, let it behave in ways the user and the market can remember.

After the app does not mean after services, after brands, after records, or after persuasion. It means after the assumption that coordination begins inside someone else's app and ends inside one assistant's private summary.

The next interface is not a perfectly neutral mind. It is a governed conversation among interested agents, with the user's assistant there to keep the thread honest.

The work is to write the rules well enough that advocacy becomes evidence, competition becomes comparison, and self-interest becomes a reason to share.

Part of the spine; the governance companion to Why Agentic Commerce Must Be a Forum. ← All essays