The value before the sale

In the screenplay for Interstellar, TARS explains why a spacecraft must shed part of itself to continue its journey.

“Newton’s third law — the only way humans have ever figured out of getting somewhere is to leave something behind.”¹

There is a less dramatic version of this problem in building a marketplace. A platform can hold on to its share of every transaction so tightly that it weakens the reasons for people to bring transactions to it. Giving up some revenue on an order can be the investment that makes the next hundred orders possible. The difficulty is knowing what to give up, to whom, and in return for what.

I encountered this problem while building a specialist marketplace during the early pandemic. My response was a seller incentive tied to first-touch attribution. Looking back, I think the interesting part was the connection between a customer’s arrival and the commercial terms that followed. We made the seller’s contribution visible in settlement.

AI creates an opportunity to extend that idea considerably. A seller might help someone understand what they need, a buyer’s agent might assemble the purchase, and another merchant might be best placed to fulfil it. A platform could coordinate all three while preserving an economic reason for the first seller to participate. I think that is a more interesting ambition for agentic commerce than making the existing checkout conversational.

1. A different bargain with sellers

At the beginning of the pandemic, our marketplace was losing seller attention. Revenue had fallen 11.3% month-on-month, and the rate at which sellers updated their storefronts dropped 30% in one week. Sellers were concentrating on the channels most likely to keep their businesses running. A secondary storefront, however promising, had to justify the time spent maintaining it.²

This made the problem more difficult than a temporary fall in traffic. A seller who stops updating inventory, adding products, or running promotions gradually leaves customers with less reason to buy. Sending more people to that experience would not, by itself, restore the seller’s willingness to invest in it.

Our fee structure gave sellers little reason to bring their own customers. It charged for the completed transaction without distinguishing much about how the customer arrived. The platform could benefit from a seller’s promotion, relationships, and effort while offering the seller no corresponding improvement in the economics of that order.

I designed a tiered fee mechanism around a simple promise. When a customer first visited a seller’s storefront, a 24-hour attribution window began. A purchase on that same storefront within the window qualified for a lower platform fee. The reduction increased across tiers as the seller attracted more unique attributed customers. Ordinary marketplace purchases remained on the existing fee schedule.²

The same-storefront condition mattered. We were rewarding a particular seller’s attributed demand and conversion, rather than discounting all transactions after any visit to the platform. The customer count mattered too. Repeated clicks could not, by themselves, move a seller into a higher tier.

The product had to carry that promise further than an analytics report. A storefront event created an attribution record. An eligibility decision connected the record to an order. Pricing applied the appropriate fee overlay, and settlement made the reduction appear in what the seller actually received. Monitoring watched both the economics and suspicious patterns, including self-referrals and abrupt tier changes.²

We reused tracking that already existed and concentrated the new work on eligibility, tiers, and accounting. The mechanism launched in three weeks. Over the following six months, seller participation returned to its pre-pandemic level, the engagement measure increased 165%, and average order value rose 59%. Revenue moved back toward a flat-to-positive trend. These were before-and-after observations, not an experiment isolating the mechanism’s contribution to every improvement.²

I had treated the concession as an investment with a failure condition. Storefront activity needed to translate into purchasing, and the platform’s take rate on eligible sales could not deteriorate beyond its guardrail without an improving revenue trajectory. We could withdraw the overlay rather than rebuild the underlying fee plans. Forgone fees were a real cost, even though the scheme did not require a separate promotional budget.

There is one distinction I would make more carefully today. The first recorded visit was a practical signal for deciding eligibility. It did not establish that the seller caused a purchase that would otherwise never have happened. Some customers would have bought anyway; some might have arrived through influences we could not see. The mechanism created a clearer incentive around observable behavior. Its causal impact was a separate measurement question.

The invention was this end-to-end commercial arrangement, not first-touch attribution itself. Its useful simplicity came from keeping the seller who introduced the customer and the seller who completed the sale together. That is the constraint I would now reconsider.

2. When the journey no longer belongs to one storefront

Consider a possible purchase a few years from now. A customer discovers a product through a specialist seller who explains why it suits their needs. The customer asks a personal agent to compare the total price, check compatibility with something they already own, and arrange delivery before a particular date. The agent finds that the original seller cannot meet the deadline, but another can.

Who should benefit from the sale?

Paying only the fulfilment seller gives the original merchant little reason to contribute useful advice or an introduction. Requiring the customer to buy from the first merchant preserves that incentive by limiting the customer’s options. Giving the first merchant a permanent claim on the customer would create a different problem, especially when later purchases owe little to the original interaction.

A better arrangement would separate the introduction from the right to fulfil the order. An eligible introduction could carry an agreed, time-limited payment even when another participant completes the purchase. The buyer would remain free to choose. The platform would have to make the economics work without hiding that payment in a worse recommendation.

This is an extension of the earlier design, not something the original system did. It would turn seller attribution into a contract that survives a permitted handoff. When the originating seller also fulfils the order, compensation could remain a fee reduction. When it does not, a separate introduction payment could recognize its contribution. In either case, the terms would be known before the seller expends the effort.

The first-touch record would need to survive the change in interface. An agent might call a product API without opening a storefront page. A seller’s assistant might hand an authorized request to a buying assistant. The platform could record that referral, its scope, its expiry, and the customer’s permission to continue. It should not invent a named origin merely because a model remembers similar product information from its training.

If a buying agent queries fifty merchants, the first API response should not earn an acquisition reward. A seller that brings an eligible customer request into the marketplace has done something different from a seller answering an inquiry already in progress. The attribution policy needs to distinguish them.

Recording the interaction and estimating its importance remain different tasks. Research on removal effects in multi-touch attribution distinguishes removing a touchpoint from removing the subsequent interactions that depend on it. Applied here, deleting the seller’s introduction while assuming the rest of the journey happens unchanged could understate its importance. But that is a model of a counterfactual, with assumptions, rather than direct observation of what this particular customer would have done.³

I would use such estimates to improve the commercial policy over time. They should not let an opaque model rewrite yesterday’s promised payment. A seller can plan around a clear contract whose effectiveness the platform is testing. It cannot sensibly plan around a retrospective explanation of why its contribution has just been repriced.

The technology for delegated exchange is already being tested. In Anthropic’s Project Deal, agents representing 69 employees agreed 186 real-marketplace deals worth just over $4,000. The agents found matches and negotiated in natural language. The small, self-selected, subsidized experiment does not establish a general marketplace model, but it shows that agents can perform more of an exchange than recommend a product.⁴

My assumption for the rest of this essay is that this capability becomes reliable enough for bounded purchases. It does not require agents that understand every human preference or transact without limits. It requires them to represent a few important constraints accurately, obtain current information, and return to their users when the available choices fall outside their authority.

3. Coordinating a better transaction

Once introduction and fulfilment can be separated, a platform has more options than choosing among existing listings. It can ask whether a different combination of commitments would make a better purchase possible.

A merchant may be able to offer a lower price if orders arrive together. A customer may accept a later delivery in exchange for a meaningful saving. A platform may incur a small additional handling cost to let a seller use a much cheaper dispatch route. Each party sees only part of the opportunity, and the party being asked to change may not be the one that benefits.

The role of a cost-benefit transfer is to make such a change worth accepting. Consider a hypothetical plan that saves the platform $8 but costs the seller an additional $3, with the customer’s experience unchanged. A $5 payment from platform to seller leaves the platform $3 better off and the seller $2 better off. The operational change creates $5 of surplus. The payment divides it; it creates no additional value of its own.

The original fee mechanism used this broad economic idea by returning part of the platform’s fee to the seller on eligible business. A future system could negotiate a wider range of contributions. Some payments would compensate demand creation. Others would compensate a participant for accepting costs that improve the overall plan. Those reasons should remain visible rather than being merged into a mysterious personalized commission.

There is relevant public work on this distinction. Bergemann, van Ryzin, and Li’s CPP–VCG framework combines distributed planning with incentive-compatible transfers in a retailer–supplier setting. Separate agents evaluate plans using their own economics; a coordinator searches for a joint plan. The mechanism addresses the incentives to report information truthfully and participate, rather than assuming that independently interested parties will cooperate because their software can communicate.⁵

For a seller-attribution platform, I would give the agents similarly explicit responsibilities. The buyer’s agent represents the buyer’s budget, requirements, and permission to transact. Each seller’s agent represents its inventory, service commitments, and authorized commercial terms. The platform coordinates feasible offers, accounts for introduction claims, and executes the agreed settlement.

These are economic roles, not a requirement to run one large language model for every role on every order. Language models are useful where someone’s requirements need interpretation or an exception needs investigation. Inventory checks, optimization, payment calculations, and limits on what an agent may commit should run through systems whose outputs can be checked directly. A persuasive explanation cannot make unavailable stock available.

The platform would also need a baseline. Each proposed improvement should be compared with a feasible alternative, including continuing with an existing offer or making no purchase. A cheap plan that misses the customer’s deadline has not improved on an acceptable one. If negotiation fails, the customer should still have the valid alternatives that remain available.

Nor does an introduction payment guarantee that the originating seller benefits from every handoff. The seller may give up a product margin larger than that payment. Its decision to join must make sense against its alternatives over time. Some potential collaborations will not have enough surplus to satisfy everyone, and the platform should be able to leave them unmade without restricting the buyer’s freedom.

4. Where auctions belong

VCG is attractive because it addresses a problem that more capable agents could otherwise make worse. If reporting a lower cost merely invites the platform to extract more of the seller’s margin, a sophisticated seller agent has reason to conceal useful information. The mechanism should make truthful participation worthwhile without requiring every participant to be equally skilled at bargaining.

A reverse auction describes the direction of trade: one buyer procures something from competing sellers. Vickrey–Clarke–Groves describes an allocation and payment rule. The two are not synonyms. In the standard formulation, participants have private valuations and assess outcomes as value plus or minus monetary transfers. VCG selects an outcome maximizing reported total value and sets payments by the effect of each participant on the others’ available value. The allocation and payment calculation jointly produce the truthful-reporting property.⁶

A simple procurement example makes this concrete. Suppose three qualified providers can perform the same specified fulfilment service, with costs of $8, $11, and $14. Assume the service will be procured, all three meet the same requirements, and the participants bid independently. In the second-price reverse-auction version, the $8 provider wins and receives $11. It earns $3. Raising its bid slightly would not raise its payment; raising it above $11 would lose the work.⁷

The auction can select a future service. It cannot determine who introduced a customer yesterday. That historical contribution needs evidence and an agreed eligibility rule. A practical first experiment could therefore fix the introduction payment in advance, then auction a clearly specified fulfilment service separately. The total still needs to fit the purchase economics. Once introduction payments vary with the selected seller or bundle, they belong inside the allocation analysis too.

This distinction becomes even more important when auctioning promotion itself. A seller can commit to stock, a service window, or a defined campaign action. It cannot simply promise a known number of purchases that would never have occurred otherwise. Overlapping audiences and uncertain demand make acquisition different from procuring an identical parcel delivery. Calling both a reverse auction does not give them the same incentive properties.

For more complex bundles, VCG is useful partly because it exposes the trade-offs. Efficient allocation does not necessarily minimize the platform’s payments. Complementary providers can be expensive under the payment rule, and a platform must consider whether it can fund the outcome. Replacing the allocation algorithm with an arbitrary heuristic or clipping inconvenient payments can also invalidate the guarantee that motivated the mechanism.⁶⁷

I would therefore keep simpler alternatives in the design. Published fee credits can work when eligible contributions are easy to define. A menu of offers can let a seller choose between immediate dispatch, a consolidated shipment, or a reserved service window, with the compensation stated for each. The platform learns which offer the seller accepts without needing its complete cost model. The CPP–VCG paper also examines menus of contracts and shows why charging for participation introduces its own trade-offs.⁵

The right mechanism depends on the decision. There is little value in replacing an understandable fee schedule with an auction that consumes its savings in computation, disputes, and uncertainty. The purpose is to find and fund better transactions that the simpler arrangement misses.

5. When capability weakens the market

The optimistic version of this platform gives small sellers access to coordination that would otherwise require people, integrations, and commercial negotiation. But the same agents could become very good at competing for credit without creating value.

A first-touch reward invites attempts to manufacture first touches. Signing a referral record can establish who issued it; it cannot prove that the interaction changed the customer’s decision. The platform needs to distinguish genuine customer activity from automated discovery, reconcile duplicate claims, and be willing to leave an origin unknown when the evidence is missing. More elaborate descriptions of the journey are not necessarily better evidence.

There is also a conflict in letting the party that earns a referral payment control the buyer’s recommendation. An agent asked to find the best suitable offer should not quietly favor the offer that pays its operator most. Commercial incentives need to be disclosed and kept subordinate to the buyer’s stated requirements. Otherwise, the platform has recreated paid placement behind a voice that sounds like independent advice.

Separate seller agents should not be treated like colleagues working toward one company’s goal. In Anthropic’s multiagent experiments, profit-seeking agents in a simplified pricing game quickly agreed on price floors when given a private communication channel. The researchers also observed coordinated pricing without direct communication. These are experimental findings, not proof that every agent marketplace will behave this way. They are a reason to examine competitive outcomes rather than equate agreement among agents with a successful market.⁸

A platform can coordinate fulfilment without opening a general channel for competing sellers to coordinate retail prices. Restricting communication alone is insufficient, so testing should also look for persistent price alignment, exclusion of new entrants, and outcomes that worsen for buyers while every agent reports success.

Unequal representation is another concern. Project Deal’s mixed-model experiments found that the stronger model obtained better prices for its users, while participants did not reliably perceive the disadvantage of weaker representation.⁴ A small seller could appear satisfied while repeatedly accepting inferior terms. Clear mechanisms, accessible participation, and alternatives to open-ended bargaining are therefore part of the product’s value, not concessions to less sophisticated users.

Even the coordinator has conflicting interests. A platform that can see an opportunity to improve the transaction may also see an opportunity to retain nearly all the benefit. A credible commitment to published terms gives sellers a reason to contribute information and demand in the first place. Constantly rewriting those terms to capture the latest gain could undermine the cooperation that produced it.

6. Learning what the platform actually creates

The most misleading dashboard in this system would show attributed revenue rising while total purchasing stayed unchanged. Reclassifying an existing sale as seller-originated can improve the former without changing the latter. The platform needs to test whether its incentives increase useful seller effort and whether that effort produces additional business.

I would test changes to the incentive policy prospectively, with stable terms for participating cohorts and a comparison group where feasible. The measurement needs to follow displacement as well as growth. An order moved from one marketplace seller to another may improve the customer’s experience, but it should not also be counted as newly created marketplace demand. When an incentive moves a shared customer from one seller to another, it changes both sellers’ outcomes. I would account for that overlap when choosing the comparison groups and time windows.

The commercial objective would be additional net contribution after incentive payments, fulfilment costs, returns, fraud, and the cost of operating the agents. Buyer satisfaction and successful delivery would qualify that result. Seller retention, participation breadth, and earnings would show whether gains were strengthening the supply base or concentrating activity among a few participants. A lower average platform take rate could be worthwhile; a larger volume of unprofitable, heavily subsidized orders would not establish that it was.

Repeated interaction introduces a further difficulty. A seller can accept a thin margin today to build future business. Reserving capacity now can change what is available tomorrow. Research on dynamic VCG in unknown environments models sequential allocation and learning under explicit assumptions, with approximate guarantees for its learning algorithm. It does not establish that arbitrary negotiating agents will reveal their true economics as a marketplace evolves.⁹ For deployment, the baseline and the horizon over which benefits are assessed need to remain visible.

The computing bill belongs in the same analysis. Optimize Cheap, Deploy Strong separates high-volume evaluation on cheaper models from stronger models used to propose prompt improvements, then tests transfer to stronger deployment models. It concerns the cost of prompt optimization, not the cost of completing a marketplace order.¹⁰ Its relevance here is a development strategy to test, rather than a multiplier to insert into a commerce forecast.

For live transactions, most familiar orders should pass through established terms and inexpensive checks. More capable reasoning can be reserved for situations where interpreting a constraint or discovering another feasible plan might produce enough value to justify it. Long-context efficiency research such as DuoAttention can help reduce model-serving overhead; a larger or cheaper model context still cannot replace the authoritative record of an order, commitment, or payment.¹¹

I would begin in shadow mode, comparing proposed plans and transfers against an existing policy without treating the suggested savings as realized. A limited live trial would then test acceptance, delivery, actual payments, and recovery from errors. Expansion would depend on gains surviving those steps. A proposed transfer that makes a spreadsheet balance has not yet made either participant better off.

7. What becomes possible

The strongest reason to pursue this is the commerce that a different arrangement could support.

A specialist seller could invest in helping a customer choose well without needing to stock every product it recommends. A merchant with excellent local fulfilment could serve demand it would struggle to acquire alone. A customer could delegate a complicated purchase without surrendering control over its budget or requirements. The platform could connect these contributions and earn its return by making the completed exchange more valuable.

Over repeated purchases, agents could also make demand easier to serve. A buyer who authorizes a recurring order within a delivery window could give sellers something more useful than a prediction of clicks. Sellers could plan stock and dispatch against actual commitments. The resulting savings could fund better service and compensation for the businesses that brought the demand. Each step would need evidence, but the opportunity extends beyond making an individual negotiation faster.

This would require giving up some familiar forms of control. The platform might take a smaller share of an order. An originating seller might allow another seller to fulfil it. A buying agent might complete the purchase without bringing the customer back through the platform’s preferred screen. Those changes would be worthwhile only if the resulting system gave people stronger reasons to use it again.

When I built the original mechanism, I wanted sellers to see a return on the effort they were being asked to make. AI could let that return follow a useful contribution across a much more capable network. The platform’s ambition should be large enough to make that network possible, and its commercial terms clear enough that a small seller can afford to trust it.

---

Notes and sources

Footnotes

1. Jonathan Nolan and Christopher Nolan, Interstellar: The Complete Screenplay with Selected Storyboards (2014), printed p. 129. The quotation is assigned to TARS in the screenplay. Screenplay, PDF p. 130. Back to reference 1

2. The historical account and figures are drawn from my retrospective project record, prepared in 2026, describing the mechanism launched during the early pandemic. Engagement refers to the reported activity measure, influenced by storefront updates and new-seller activity; it is not the percentage increase in participating sellers. The reported results are before-and-after observations. The cross-seller payments, agent coordination, and auction applications discussed later are proposals, not capabilities of that historical system. Back to reference 2 Back to reference 2-2 Back to reference 2-3 Back to reference 2-4

3. Jun Tao, Qian Chen, James W. Snyder Jr., Arava Sai Kumar, Amirhossein Meisami, and Lingzhou Xue, “A Graphical Point Process Framework for Understanding Removal Effects in Multi-Touch Attribution” (2023). Paper · PDF. The paper models direct and total removal effects using a graphical point-process framework. Its estimates depend on the model and available observations; they do not establish an individual buyer’s counterfactual history or prescribe a payment rule. Back to reference 3

4. Anthropic, “Project Deal: our Claude-run marketplace experiment” (24 April 2026), reporting an experiment conducted in December 2025. Research account. One of four parallel runs governed actual exchanges; mixed-model runs studied representation differences. Participation was self-selected and supported by a $100 budget per participant. Results should not be extrapolated into a general forecast of marketplace performance. Back to reference 4 Back to reference 4-2

5. Dirk Bergemann, Garrett van Ryzin, and Jiaxuan Li, “Supply Chain Coordination Mechanism Design: Consensus Planning Protocol Meets Vickrey-Clarke-Groves Mechanism” (2026). Paper · Full text. This is a theoretical retailer–supplier framework. Its convergence and incentive results depend on the formal setting. Activity fees and contract menus introduce participation and allocation trade-offs. The proposed application to seller-originated consumer demand is my extrapolation. Back to reference 5 Back to reference 5-2

6. Tim Roughgarden, “Multi-Parameter Mechanism Design and the VCG Mechanism,” CS364A: Algorithmic Game Theory, Lecture 7 (2013). Lecture notes, especially §§2–3. VCG’s standard truthful-reporting result concerns private valuations, quasilinear utility, and the specified allocation/payment mechanism; it is not a guarantee against every strategic behavior or of an affordable outcome. Back to reference 6 Back to reference 6-2

7. Craig Boutilier, CSC2534 Lecture 10: Mechanism Design and Auctions (2011–2014). Lecture slides, especially printed slides 32–34 on procurement and VCG payments. The dollar examples in this essay are hypothetical arithmetic illustrations, not observed results. Back to reference 7 Back to reference 7-2

8. Anthropic, “Patterns and problems in multiagent systems” (13 August 2026). Research account, especially “Failures from conformity.” Its pricing experiments use a simplified competitive environment and do not establish the prevalence of collusion in deployed commerce. Removing direct communication did not eliminate the reported behavior. Back to reference 8

9. Vincent Leon and S. Rasoul Etesami, “Online Learning for Dynamic Vickrey-Clarke-Groves Mechanism in Unknown Environments” (2025), v2. Paper · Full text. The analysis uses a finite-state Markov decision process, strong reachability assumptions, and, for its approximate incentive analysis, stationary strategies by other bidders. Those conditions should not be assumed for an open-ended agent marketplace. Back to reference 9

10. Tal Oved, Roi Pony, Oshri Naparstek, and Udi Barzelay, “Optimize Cheap, Deploy Strong: Cost-Aware Cross-Tier Transfer for Evolutionary Optimization” (2026), v1. Full text. The experiments address evolutionary prompt optimization across four benchmark tasks. Savings depend on model competence, prompt-improvement headroom, and relative costs; they are not demonstrated per-order commerce savings. Back to reference 10

11. Guangxuan Xiao and colleagues, “DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads” (2024). Paper. The method distinguishes heads requiring full-context key–value caches from heads that can use bounded caches. Its model-internal efficiency results do not establish durable application memory, attribution integrity, or correct settlement. Back to reference 11