Capability-Priced Micro-Markets (CPMM)

• CPMM is defined as a framework where autonomous agents trade decision-critical capabilities and verified information through micropayments and cryptographically assured exchanges.
• It integrates mechanisms of price discovery, sequential negotiation, and trust via reputational scoring and cryptographic attestation to ensure high-quality transactions.
• The model applies decentralized protocols and cost-aware, iterative information acquisition to optimize secure and efficient agent-based e-commerce.

Searching arXiv for the specified CPMM papers to ground the article in the cited sources. Capability-Priced Micro-Markets (CPMM) is a micro-economic framework for agentic commerce in which autonomous agents buy and sell narrowly scoped capabilities or verified information items through micropayments, negotiated commitments, and cryptographically attested exchange. In one formulation, CPMM is presented as a market for verified product information in agentic e-commerce, where buyer agents spend fractions of a cent to progressively unlock seller- and reviewer-supplied data under a freemium model, with reviewer trust scored reputationally (Ventirozos et al., 23 Jun 2026). In another, CPMM is framed as a general framework for the agentic web over HTTP 402, integrating capability-based security, micropayment rails, and structured negotiation into a unified system for decentralized agent ecosystems (Huang et al., 4 Mar 2026). Across both formulations, the central shift is from ranking and recommendation toward priced access to decision-relevant, verifiable evidence.

1. Formal definition and market primitives

A CPMM market can be defined as $M=(C,B,S,P)$, where $C=\{c_1,\dots,c_n\}$ is the universe of capabilities or data items, $B$ is the buyer-agent with private valuation function $v_b:C\to\mathbb{R}_+$ and hard budget $B_{\max}$, $S$ is the seller side, and $P:C\to\mathbb{R}_+$ is the posted price function (Ventirozos et al., 23 Jun 2026). When the buyer acquires a set $C'\subseteq C$, it pays

$R_s(C')=\sum_{c\in C'}P(c),$

and its net utility is

$U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$

A seller’s revenue from a listing $C=\{c_1,\dots,c_n\}$0, with disclosed bundle $C=\{c_1,\dots,c_n\}$1, is $C=\{c_1,\dots,c_n\}$2 (Ventirozos et al., 23 Jun 2026).

The data items in $C=\{c_1,\dots,c_n\}$3 are explicitly decision-relevant units such as “service-history digest,” “finance check,” and “fitment confirmation” (Ventirozos et al., 23 Jun 2026). A running example describes a used-car setting in which $C=\{c_1,\dots,c_n\}$4 includes $C=\{c_1,\dots,c_n\}$5 “mileage check,” $C=\{c_1,\dots,c_n\}$6 “service-history digest,” and $C=\{c_1,\dots,c_n\}$7 “independent condition report,” with prices $C=\{c_1,\dots,c_n\}$8, $C=\{c_1,\dots,c_n\}$9, and $B$0. If the buyer acquires $B$1, then $B$2 and $B$3 (Ventirozos et al., 23 Jun 2026).

A broader formulation models CPMM interactions as a repeated bilateral game with incomplete information. In each period $B$4, the buyer requests quality $B$5 and maximum price $B$6, and the seller quotes price $B$7 or rejects via HTTP 402. If trade occurs at $B$8, buyer payoff is $B$9 and seller payoff is $v_b:C\to\mathbb{R}_+$0; otherwise $v_b:C\to\mathbb{R}_+$1 (Huang et al., 4 Mar 2026). Buyer and seller maintain beliefs $v_b:C\to\mathbb{R}_+$2 and $v_b:C\to\mathbb{R}_+$3 over private types and update them through Bayes’ rule after observing quotes or rejection (Huang et al., 4 Mar 2026).

This pairing of item-level utility accounting with repeated-game learning suggests that CPMM operates simultaneously as a micro-pricing scheme and as a mechanism for sequential information revelation. A plausible implication is that the framework is intended to cover both discrete evidence purchases in e-commerce and more general service transactions among autonomous agents.

2. System architecture and protocol stack

One CPMM architecture consists of four components: User $v_b:C\to\mathbb{R}_+$4 Buyer-Agent, Buyer-Agent $v_b:C\to\mathbb{R}_+$5 Marketplace Engine, Marketplace Engine $v_b:C\to\mathbb{R}_+$6 Supply, and Supply $v_b:C\to\mathbb{R}_+$7 Buyer-Agent (Ventirozos et al., 23 Jun 2026). In this arrangement, the user provides intent, budget, and constraints; the buyer-agent interacts with the marketplace engine through x402/AP2 micro-payments; the marketplace engine sends evidence requests to supply-side actors; and the supply side returns verified data payloads (Ventirozos et al., 23 Jun 2026). A stylized request protocol for a single datum $v_b:C\to\mathbb{R}_+$8 proceeds by sending a request, receiving a price quote, authorizing an AP2 token if the price is within the remaining budget, paying via x402, and then receiving {type:"Data", cap:c, payload:data_c, proof:\sigma} (Ventirozos et al., 23 Jun 2026).

A more general protocol stack synthesizes three foundational technologies: NANDA, HTTP 402 / X402 / H402 micropayments, and ACNBP (Huang et al., 4 Mar 2026). NANDA provides a capability-based security model in which agents hold unforgeable tokens

$v_b:C\to\mathbb{R}_+$9

an Agent Name Service (ANS) for federated, privacy-tuned registry of agent capabilities, and attestation and delegation chains for cryptographic proof of issuance (Huang et al., 4 Mar 2026). HTTP 402 is enriched with headers including X402-Payment-Required, H402-Payment-Key, and H402-Payment-Signature, using ephemeral keys (Ed25519) and Schnorr signatures for atomic, low-overhead micropayment commits, with batch settlement and conditional pay-on-delivery semantics (Huang et al., 4 Mar 2026). ACNBP contributes a ten-step Agent Capability Negotiation & Binding Protocol: Discover $B_{\max}$0 Pre-screen $B_{\max}$1 Negotiate-Request $B_{\max}$2 Negotiate-Response $B_{\max}$3 Bind $B_{\max}$4 Commit $B_{\max}$5 Execute $B_{\max}$6 Verify $B_{\max}$7 Release $B_{\max}$8 Audit (Huang et al., 4 Mar 2026).

The integration is explicit: NANDA supplies verifiable discovery and access tokens, ACNBP carries semantic and economic proposals such as pricing models, SLAs, and payment terms, and HTTP 402 / X402 / H402 provides micropayment rails tied to quality attestations generated under NANDA guarantees (Huang et al., 4 Mar 2026). In the e-commerce formulation, the marketplace additionally performs entity resolution to route a request to exactly one seller or data vendor, and AP2 binds the payment to the principal’s budget (Ventirozos et al., 23 Jun 2026).

3. Negotiation, pricing, and equilibrium structure

CPMM does not require prices to remain fixed. A dynamic version models data pricing and bundle formation as a two-player alternating-offers game in the Rubinstein style (Ventirozos et al., 23 Jun 2026). The buyer has valuation $B_{\max}$9 for any bundle, the seller has cost $S$0 to reveal data, and the proposal space is any $S$1 pair consisting of a bundle and a total price. The buyer offers $S$2, the seller can accept, reject, or counter with $S$3, and if a deal occurs at price $S$4, the payoffs are $S$5 for the buyer and $S$6 for the seller (Ventirozos et al., 23 Jun 2026). With common discount factor $S$7, the subgame-perfect equilibrium price partition satisfies $S$8 in the limit $S$9 (Ventirozos et al., 23 Jun 2026).

The broader CPMM paper frames price discovery through repeated bilateral interaction under incomplete information and argues that the mechanism converges to a constrained Radner equilibrium (Huang et al., 4 Mar 2026). A tuple $P:C\to\mathbb{R}_+$0 is a constrained Radner equilibrium if, given beliefs $P:C\to\mathbb{R}_+$1, the pair $P:C\to\mathbb{R}_+$2 solves each agent’s static best-response problem, $P:C\to\mathbb{R}_+$3 is consistent with on-equilibrium behavior, and quality $P:C\to\mathbb{R}_+$4 lies in each agent’s feasible set (Huang et al., 4 Mar 2026). The equilibrium conditions are presented as a price–quantity fixed point: $P:C\to\mathbb{R}_+$5 The proof outline states that Bayesian learning is shown to be a contraction in the space of beliefs under mild Lipschitz conditions, that belief updates are modeled as a stochastic approximation with mean-field fixed point $P:C\to\mathbb{R}_+$6, that martingale convergence implies $P:C\to\mathbb{R}_+$7 almost surely at rate $P:C\to\mathbb{R}_+$8, and that best-response mappings then converge to the Nash fixed point $P:C\to\mathbb{R}_+$9 of the stage game (Huang et al., 4 Mar 2026).

In the e-commerce paper, buyer and seller agents are said to exchange offers over bundles of evidence and to learn from when the buyer walked away in order to re-price future items (Ventirozos et al., 23 Jun 2026). This suggests a micro-market in which screening, negotiation, and adaptive repricing are part of the core market design rather than auxiliary mechanisms.

4. Cost-optimal information acquisition

A central CPMM problem is sequential, budget-constrained information acquisition. The buyer’s planning problem is described as “Cost-aware tool use,” where the objective is to maximize expected net utility subject to budget and stopping constraints (Ventirozos et al., 23 Jun 2026). Let $C'\subseteq C$0 be the set of available queries, and let each $C'\subseteq C$1 return a data item $C'\subseteq C$2 with stochastic informativeness $C'\subseteq C$3, such as reduction in purchase-risk (Ventirozos et al., 23 Jun 2026). The problem can be formulated as

$C'\subseteq C$4

subject to

$C'\subseteq C$5

or equivalently as maximizing

$C'\subseteq C$6

where $C'\subseteq C$7 trades off information gain against cost (Ventirozos et al., 23 Jun 2026).

The same source states that solving this problem is a POMDP in which the buyer’s belief state is over unseen risks and actions are “buy datum $C'\subseteq C$8 at cost $C'\subseteq C$9” (Ventirozos et al., 23 Jun 2026). Approximate solutions use greedy expected-value-of-information / cost heuristics (Ventirozos et al., 23 Jun 2026). This places CPMM squarely within decision-theoretic tool use rather than conventional chatbot interaction. The explicit claim is that cost-optimal information acquisition, data pricing and negotiation, real-time entity resolution, grounded value exchange, and privacy-preserving persona modelling deserve the field’s attention, rather than chat fluency (Ventirozos et al., 23 Jun 2026).

An illustrative sequence in the general framework shows this process for service procurement: discovery through NANDA/ANS, ACNBP Negotiate-Request with a quality requirement such as {accuracy:0.98, latency:200ms}, a maximum price, and a disclosure parameter $R_s(C')=\sum_{c\in C'}P(c),$0; an ACNBP Negotiate-Response quoting price and SLA; binding and commitment; an HTTP 402 payment with H402-Payment-Key, H402-Payment-Amount, and H402-Payment-Signature; then seller-side verification, delivery, and issuance of a TEE-based quality attestation before escrow release (Huang et al., 4 Mar 2026). Although this example concerns a service rather than product evidence, it presents the same underlying logic of staged acquisition under explicit price and quality constraints.

5. Trust, attestation, and privacy elasticity

CPMM introduces multiple trust layers. In the e-commerce formulation, third-party reviewers stake a bond $R_s(C')=\sum_{c\in C'}P(c),$1 and earn payments for issuing opinions $R_s(C')=\sum_{c\in C'}P(c),$2 (Ventirozos et al., 23 Jun 2026). Their trust score $R_s(C')=\sum_{c\in C'}P(c),$3 is updated according to

$R_s(C')=\sum_{c\in C'}P(c),$4

where $R_s(C')=\sum_{c\in C'}P(c),$5 if the opinion is validated by future outcomes and $R_s(C')=\sum_{c\in C'}P(c),$6 otherwise (Ventirozos et al., 23 Jun 2026). A reviewer’s price multiplier can then be set as

$R_s(C')=\sum_{c\in C'}P(c),$7

Misleading reviews cause slashing of $R_s(C')=\sum_{c\in C'}P(c),$8 and a drop in $R_s(C')=\sum_{c\in C'}P(c),$9 (Ventirozos et al., 23 Jun 2026). The paper characterizes this as a reputation or trust score settled by staking (Ventirozos et al., 23 Jun 2026).

In the general framework, trust is also mediated by cryptographic attestation. Capability forgery resistance is defined in terms of breaking the signature scheme, H402 nonces and Schnorr commitments are used to prevent replay and double spend, quality attestation integrity is tied to TEE attestation proofs such as SGX or TrustZone, and Sybil resistance is described by issuance cost that grows with identity count while economic benefit is sublinear in the number of Sybil identities (Huang et al., 4 Mar 2026). All messages are integrity-protected and non-repudiable, with logs suitable for automatic audit and reputation update (Huang et al., 4 Mar 2026).

A distinct theoretical contribution is “privacy elasticity of demand” (Huang et al., 4 Mar 2026). Let $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$0 denote the normalized disclosure level and let $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$1 be expected transaction quantity or willingness to trade. Then privacy elasticity is defined as

$U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$2

It measures the percentage change in demand for an agent’s services arising from a one-percent change in information disclosure (Huang et al., 4 Mar 2026). If $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$3, more transparency increases trade; if $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$4, disclosure may reduce demand by increasing competitive risk (Huang et al., 4 Mar 2026). Agents choose $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$5 by equating marginal gain in expected revenue with marginal privacy-cost: $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$6 This formalizes a trade-off that the e-commerce paper presents operationally as keeping the buyer persona local, sharing only minimal derived signals such as “already spent £5,” and using differential privacy or encrypted transcripts for human oversight, with the persona remaining user-held, portable, and inspectable (Ventirozos et al., 23 Jun 2026).

6. Entity resolution, grounding, and persona handling

Entity resolution and knowledge-base construction are treated as core challenges in CPMM (Ventirozos et al., 23 Jun 2026). The requirements include on-the-fly linking of heterogeneous attribute labels, such as determining whether “one-owner history” corresponds to a particular canonical capability, and grounding protocols in which buyer and seller run a short consensus dialogue to verify that they refer to the same object or attribute (Ventirozos et al., 23 Jun 2026). Agents maintain an ontology $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$7 for each listing $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$8, and incoming requests such as service_history are mapped to a canonical slot like $U_b(C')=\sum_{c\in C'}v_b(c)-R_s(C').$9 (Ventirozos et al., 23 Jun 2026). Grounding uses a simple triple-exchange:

• Buyer $C=\{c_1,\dots,c_n\}$00 Seller: “I seek attribute $C=\{c_1,\dots,c_n\}$01.”
• Seller $C=\{c_1,\dots,c_n\}$02 Buyer: “$C=\{c_1,\dots,c_n\}$03 means {schema path …}.”
• Buyer: “Confirmed.” These steps are presented as technical mechanisms for grounding and schema alignment rather than as optional metadata conventions (Ventirozos et al., 23 Jun 2026).

The general CPMM architecture supplies complementary infrastructure through NANDA and ACNBP. ANS functions as a federated, privacy-tuned registry of agent capabilities, while ACNBP provides typed negotiation and binding stages with extension slots for CPMM economic payloads (Huang et al., 4 Mar 2026). This suggests that entity resolution in product-centric settings and capability discovery in service-centric settings are structurally related: both are concerned with mapping local semantic descriptions into cryptographically and economically actionable references.

Persona handling is similarly constrained by privacy requirements. The e-commerce sketch requires keeping the persona $C=\{c_1,\dots,c_n\}$04, including budget, preferences, and risk thresholds, in a local enclave and sharing only minimal derived signals (Ventirozos et al., 23 Jun 2026). The phrase “serious-buyer” signals appears specifically in this context (Ventirozos et al., 23 Jun 2026). Differentially private or encrypted transcripts are proposed for human oversight, and the ethical requirement is that the persona should be user-held, portable, and inspectable (Ventirozos et al., 23 Jun 2026). A plausible implication is that CPMM treats buyer modeling not as a platform-owned profiling resource but as a user-controlled state that conditions negotiation and search.

7. Mechanism claims, scope, and open directions

The e-commerce paper characterizes CPMM as a truthful, quality-rewarding mechanism (Ventirozos et al., 23 Jun 2026). The incentive logic is stated as follows: high-quality products expose more verifiable data at moderate cost, causing buyers to pay less risk premium and proceed further down the data trail, which generates more revenue; low-quality sellers either charge high prices and lose buyers early or remain opaque and deter serious buyers; third-party reviewers stake reputation and face economic slashing if contradicted by later outcomes; and hard constraints such as “must have clean title” cause buyers to walk away swiftly on any violation, signaling true buyer preferences and preventing wasteful negotiations (Ventirozos et al., 23 Jun 2026). The paper further states that offers surviving buyer screening reflect genuine product quality rather than SEO copy or paid ranking, and that because every datum is cryptographically attested and priced on its marginal contribution to decision utility, the mechanism enforces a form of “information-theoretic competition” in which goods compete by revealing real, verifiable capabilities $C=\{c_1,\dots,c_n\}$05 (Ventirozos et al., 23 Jun 2026).

The general framework extends CPMM beyond e-commerce into autonomous agent commerce on the web. It describes scalability through federated ANS with consistent hashing, asynchronous quality-verification workers, batching and caching of economic payloads, and off-chain settlement with micropayment aggregation (Huang et al., 4 Mar 2026). It also lists potential extensions: combinatorial auctions for complex workflows with VCG payments under budget constraints, cross-chain atomic swaps for multi-currency micropayment support, advanced privacy using zk-SNARK proofs of capability and selective disclosure, and formal verification of the ACNBP+CPMM state machine in a tool like TLA+ (Huang et al., 4 Mar 2026).

A common misconception would be to reduce CPMM to a payment wrapper around existing recommendation systems. The available descriptions do not support that reading. Instead, CPMM is defined by the joint use of capability discovery, negotiation, micropayment-triggered access, belief updating, attestable delivery, and explicit utility or payoff formulations (Huang et al., 4 Mar 2026, Ventirozos et al., 23 Jun 2026). Another misconception would be to treat CPMM as only a seller-pricing scheme. The sources instead present it as a bilateral mechanism in which buyer budgets, private valuations, disclosure choices, and stopping policies are equally central (Ventirozos et al., 23 Jun 2026). Taken together, these formulations place CPMM at the intersection of mechanism design, secure protocol engineering, and sequential decision-making for autonomous agents.