
Machine-Native Trading: Current State and Missing Infrastructure
Author: Waterdrip Capital
Introduction: Agents Are Gaining a Constrained "Discretionary Budget"
Over the past two years, agent capabilities have evolved rapidly. Early large language models primarily handled information generation, where users posed questions and models returned text. Tool calling then enabled models to search the web, query databases, execute code, and interact with software. Taking it a step further, agents began breaking down goals, formulating plans, and adjusting their actions across multiple execution rounds based on external feedback.
When execution targets are limited to free tools or internal enterprise systems, call permissions can be pre-configured by developers. But high-quality capabilities in open markets typically require payment: real-time financial data is priced per request, web scraping consumes quotas, inference and GPU compute are billed by usage, and video generation along with specialized databases carry clear price tags. For an agent to complete a task independently, it inevitably becomes a buyer during runtime.
Traditional API business models were not designed for this type of buyer. They require a human to first visit a website, register an account, link a bank card, choose a plan, safeguard the API key, and then inject that key into the program environment. Purchase decisions and actual calls are split across two different time points: humans complete procurement before the task occurs, while the software merely consumes already purchased quotas.
An agent, however, may only discover what it needs at some point during task execution. It cannot predict in advance which data source it will ultimately call, nor should it require users to create separate accounts for every potential service. Its procurement features immediacy, low transaction values, multiple merchants, high frequency, and result orientation. For an agent, the most natural experience is not "subscribe first, then call," but rather "discover the service, receive a quote, authorize payment, and obtain the result."
Agent Payment, therefore, is not about adding a payment button to a chatbot. It means software begins to possess constrained spending authority and establishes a procurement workflow belonging to machines. Humans set goals, budgets, and risk boundaries, while the agent allocates funds within those limits. Payments thus transform from a mere settlement action into a core component of the agent’s decision-making system.
1. Why Agent Payment Has Emerged as a Standalone Sector
1.1 From Tool Calling to Economic Action
The difference between agents and standard automation scripts lies in more than just reasoning capabilities. Scripts execute predefined workflows where required resources and vendors are typically hardcoded; agents, conversely, choose paths dynamically based on their environment. In a single research task, an agent might first purchase search results, decide whether an industry database is needed based on those outputs, and finally invoke another model for cross-validation. Each procurement step alters subsequent decision-making.
This "execute-and-procure" paradigm brings economic choices directly into software runtime. An agent must not only assess tool availability but also evaluate purchasing viability: does the price exceed the budget, does response latency meet task requirements, is historical fulfillment reliable, and are alternative services better suited? Traditional tool routing focuses on capability matching, whereas machine procurement simultaneously handles pricing and counterparty risk. In these transactions, the agent itself bears the risk: payment may succeed while service delivery fails.
Therefore, the core need for Agent Payment is not unconditional automated payment, but rather the controlled delegation of purchasing authority to software. Users are reluctant to hand over entire digital wallets to agents, but they are willing to allocate a few dollars for a specific task, permitting several microtransactions. While large-scale authorizations may still require long-term trust-building, smaller delegated budgets already deliver tangible value.
1.2 Micro-amounts, High Frequency, and Multi-Merchant Dynamics Reshape Payment Economics
Human internet payment infrastructure excels at handling relatively low-frequency, higher-value transactions. Credit card networks, payment gateways, and subscription systems incur fixed costs, leading merchants to bundle numerous calls into monthly plans. For API requests worth mere cents per invocation, traditional payment fees, chargeback risks, and account maintenance costs can easily outweigh the product's actual value.
Machine consumption operates inversely. An agent can initiate multiple procurements across various merchants within minutes to fulfill a single deliverable. While individual amounts are low, the high invocation frequency means transaction volume can vastly exceed human consumer levels. Stablecoins and on-chain programmable settlement provide a new economic foundation for such scenarios: capital flows continuously, payment authorization can be signed by software, and services can be directly priced per invocation.
More importantly, multi-merchant procurement changes how the API market competes. Subscription models encourage long-term vendor lock-in, whereas pay-per-use allows agents to make dynamic selections for each task. Service providers now compete not only for annual contracts but also for instantaneous demands. Pricing, performance, and fulfillment records can all impact routing outcomes in real time.
1.3 Stablecoins Are Evolving from Trading Mediums to Settlement Infrastructure
Early demand for stablecoins in crypto markets stemmed primarily from trading liquidity and capital hedging. As issuance, custody, compliance, and cross-chain infrastructure gradually mature, stablecoins are entering cross-border settlements, corporate treasury management, and internet-native payments. For machine-to-machine payments, stablecoins offer a distinct advantage: they function simultaneously as currency and as programmable digital assets that software can directly manipulate.
Card payments rely on cardholder identity, bank accounts, and regional banking networks. Agents lack legal personhood and cannot independently navigate traditional account-opening processes. A policy-constrained wallet, however, can serve as an agent's financial interface: operators fund it with limited balances, set per-transaction and session caps, and retain freeze/revoke permissions; the agent simply signs payments within its authorized scope.
This does not imply that on-chain payments are inherently superior to all traditional methods. Consumer protection, refund mechanisms, privacy, key management, and regulatory liability still require resolution. However, for machine-to-machine, low-value pay-per-call, and global service procurement, programmable stablecoins demonstrate clear fit. They finally create the opportunity to compress both "calling interfaces" and "payment interfaces" into a single network interaction.
2. Payment Rails Have Already Emerged: x402, MPP, and Native HTTP Transactions
2.1 Transforming 402 from an HTTP Status Code into a Commercial Interface
HTTP has long reserved the 402 Payment Required status code, yet it failed to establish a universal workflow for nearly 30 years. Machine payment protocols have reactivated this semantics: a client requests a paid endpoint, the server responds with a 402 and machine-readable payment terms; the client selects an acceptable plan, completes the signature or payment, and retries the request with proof of purchase.
The significance of this process lies in eliminating human registration pages. Price discovery, payment requirements, and content delivery all occur at the protocol level understandable by programs. For developers, paid APIs no longer need to build full SaaS portals around accounts, tiers, and keys; for agents, services can be discovered like standard web pages and purchased on demand.
x402 stands out as one of the most prominent open protocols on this path. It structures payment challenges and proofs around HTTP 402, enabling service providers to bill per request. MPP approaches the problem from a different ecosystem, exploring machine-oriented methods like charges and sessions. Despite differing designs, both validate a shared direction: machine payments can become integral to application protocols rather than requiring manual settlement processes built outside them.
2.2 The Long-Term Rationale of Diverse Payment Rails
The industry often anticipates a future where only one standard protocol, one settlement network, and one payment solution remain. However, from a merchant perspective, diversity holds long-term rationale. One-off data queries suit pay-per-request pricing, while continuous inference or streaming services lean toward session billing. High-value services require robust guarantees and dispute resolution, whereas low-value calls prioritize speed and cost. Different regions and enterprises will also adopt varying compliance and settlement networks.
Protocol layers will continue innovating. Merchants may adopt direct debit, pre-authentication, custodial arrangements, streaming payments, or batch settlements; networks will inevitably trade off between cost, finality, liquidity, and ecosystem tooling. For sellers, this represents free choice. For buyers, each new combination introduces an additional integration surface.
The configuration matrix in the figure below captures a cross-section of this diversity: protocols and payment solutions form the columns, chains form the rows, and each selection represents a distinct payment configuration requiring standalone integration—a table that continues to widen.

Figure 1: Payment Rail Configuration Matrix Under Fragmentation
Consequently, fragmentation does not necessarily vanish naturally as markets mature. The credit card industry never consolidated into a single network despite decades of development, nor did cloud computing converge to a single vendor. Mature markets typically do not erase differences but instead build aggregation, routing, and clearing layers atop them. Agent Payment will likely follow a similar evolutionary trajectory. This divergence is already measurable. Data from two public explorers (x402scan and mppscan) over the last 30 days (as of September 3, 2026) shows: the MPP protocol hosts 65,591 active buying wallets on the Tempo chain, x402 hosts 19,472 on Base, and only 365 wallets appear on both rails—accounting for less than 0.6% of MPP buyers and 2% of x402 Base buyers. Of these overlapping wallets, only 112 executed more than ten trades on each rail, with a significant portion being dual-rail aggregators paying via a shared key rather than buyers independently adopting a second payment track. Buyers are not flowing across rails; each track is accumulating its own independent buyer cohort.
2.3 Seller Onboarding Represents Only Half the Equation
Payment protocols first lower the barrier for merchants to receive funds. Any endpoint capable of publishing quotes, verifying proofs, and delivering services meets the basic criteria for machine-operated commerce. Consequently, more developer tools, data services, and content interfaces are transitioning into machine-purchasable states.
However, payable supply does not guarantee automatic demand. Merchants solve "how do I get paid by machines," while agents still must answer "who should I buy from, how should I pay, and how do I verify delivery afterward?" If every buyer must separately integrate each protocol, provision funding across disparate networks, and maintain independent ledgers, machine payments will simply replicate the complexity of early API integrations, substituting API keys with wallets and protocol adapters.
True adoption rates depend on the total friction of the transaction, not merely the friction of the final settlement step.
3. The Industry's Real Bottleneck: Lack of a Closed-Loop Transaction Cycle

Figure 2: End-to-End Workflow of a Machine Procurement
3.1 Milestone One: Discovering Purchasable Services
Agents require machine-readable service catalogs. A functional directory cannot merely list names and URLs; it must also describe endpoint capabilities, inputs/outputs, pricing units, supported protocols, latency, geographic restrictions, and update statuses. Mapping between natural language intent and API parameters is equally essential; otherwise, an agent knowing it "needs macroeconomic data" cannot determine which endpoint fulfills the requirement.
Open-market directories also face issues of duplication, obsolescence, and false claims. Any merchant can assert they provide high-quality data, yet agents lack the luxury of conducting days-long background checks like human procurement officers. The discovery layer must continuously verify whether endpoints are callable, quotes are accurate, and descriptions match actual return payloads.
This makes service discovery fundamentally different from traditional search. Search engines optimize for informational relevance, whereas machine procurement directories must optimize for tradability: capability alignment, price acceptability, payment compatibility, and merchant fulfillment capacity.
3.2 Milestone Two: Understanding and Comparing Quotes
Superficially, comparable APIs can all be priced per request, but in reality, quote comparability remains weak. One provider bills per request, another per output record; one bundles model inference into its base price, another charges extra; still others dynamically bill based on input length, runtime, or successful outcomes.
Agents cannot simply select the endpoint with the lowest nominal price. They must factor in total cost, delivery probability, latency, and result quality. If a cheap interface fails repeatedly, retry costs and task delays could easily inflate its effective price. Quotes must therefore be evaluated alongside service-level agreements, historical performance, and task context.
Machine-readable quotes also require explicit expiration windows and final amounts. In dynamic pricing environments, the commitment an agent signs must be definitive, not a vague price range. Operators similarly need visibility into fee composition—including service fees, network costs, and routing charges—to establish credible budgets.
3.3 Milestone Three: Capital Distribution and Cross-Rail Liquidity
If an agent must purchase services across multiple chains and protocols simultaneously, the most straightforward approach is to pre-fund balances on each network. However, this fractures small amounts of capital into numerous fragments. Idle funds sit dormant on unused networks while hotspots run dry; replenishing balances entails bridging, swapping, Gas management, and security overhead.
This complexity overwhelms individual users. For enterprises managing fleets of agents, the issue compounds: how much balance should each agent hold, who is responsible for top-ups, how to prevent erroneous fund consumption, and how to consolidate assets and expenses across disparate networks? Without a unified treasury layer, more payment rails paradoxically increase financial complexity.
Ideally, an agent views a single discretionary budget rather than scattered network balances. The underlying system handles settlement path selection, liquidity management, and transparent quoting. This principle mirrors travelers using a single card abroad: users care about total limits and exchange rates, avoiding the hassle of opening local accounts for every destination.
3.4 Milestone Four: Policy-Based Authorization
The primary concern surrounding autonomous payments is the risk of agents spiraling into uncontrolled spending. The solution is not a binary choice between total prohibition and unrestricted authorization, but the implementation of layered policies.
Per-transaction caps limit damage from single errors, session budgets constrain overall task expenditure, merchant whitelists/blacklists manage counterparties, category rules restrict purchasable content types, and rate limits prevent anomalous burst calls. High-risk or high-value transactions can trigger human confirmation. Operators define these strategies; agents operate strictly within bounds and cannot autonomously raise limits.
Wallets should also transcend mere signature functions. They must integrate with task context, identity, and audit trails to answer "which agent approved this payment, for what task, and under which policy?" Otherwise, enterprises are left with nothing but a string of on-chain transaction hashes, failing to meet internal control and cost attribution requirements.
3.5 Milestone Five: Successful Settlement Does Not Guarantee Service Delivery
Blockchains excel at proving funds moved from one address to another, but cannot inherently verify that an API returned correct content. A transaction may settle successfully while the server times out, returns error codes, or delivers data inconsistent with its marketing. For agents, this is not a marginal edge case but the core of procurement risk.
Traditional e-commerce bridges payment and delivery through logistics, reviews, and refunds. Machine services lack physical shipping; delivery may simply be a fleeting HTTP response. If payment systems only log fund movements and merchants only archive their own responses, the market lacks a unified fulfillment view spanning merchants and protocols.
Care must be taken: logging a response does not equate to proving quality. Yet linking payments to responses at least distinguishes fundamental states like "paid & delivered," "paid & failed," and "unsettled." This forms the foundational layer of facts for building machine transaction credibility.
3.6 Milestone Six: Unified Reconciliation and Liability Allocation
A single task may encompass over a dozen micro-procurements. If transactions scatter across disparate wallets, protocols, and merchant dashboards, users struggle to understand why a final deliverable incurred these specific costs. Enterprises also need to allocate expenditures to projects, teams, clients, and cost centers while retaining auditable evidence.
A unified ledger should concurrently record procurement intent, merchant details, quotes, authorization policies, settlement outcomes, response statuses, and failure reasons. It serves both financial reporting and agent optimization. Systems can analyze which data sources frequently fail, which routing paths carry higher costs, and typical procurement combinations for specific task categories.
Once payment is embedded within the reasoning chain, cost becomes a feedback signal for model decisions. Without unified reconciliation, agents optimize answers but fail to optimize the economic process of acquiring them. A significant portion of Agent Payment's long-term value stems precisely from this observability.
4. From Payment Protocols to Machine Procurement Layers
4.1 The Core Abstraction of Tomorrow Is Not "Pay," But "Buy"
Payment is an action taken after specifying an object and price; procurement covers the entire process from demand identification to acceptance verification. Exposing a `pay()` function to an agent merely enables transfers to known addresses. Exposing a `buy()` capability, however, means the system can ingest requirements, discover services, compare options, execute payments, and return verifiable results.
This distinction dictates industry specialization. Protocols provide standardized payment messages, wallets manage signatures and assets, settlement networks move value, catalogs aggregate supply, and the procurement layer orchestrates these components into cohesive tasks. While every single component matters, none alone represents a complete transaction.
Machine procurement layers must remain open. They should not mandate universal protocol migration nor dictate purchasability through closed catalogs. A more sustainable model involves supporting diverse payment rails, disclosing routing costs within quotes, and granting agents autonomous selection based on defined policies.
4.2 Buyer Aggregation May Prove More Critical Than Seller Aggregation
Internet platforms traditionally aggregate supply first, then attract consumers. In machine markets, supply already exists widely in API form; what is missing is standardized buyer infrastructure capable of sustained purchasing. An equipped agent can transform fragmented, sporadic demands into stable transaction flows.
Buyer aggregation also enhances visibility for long-tail services. Human developers default to familiar brands due to the high time cost of evaluating new suppliers; agents capable of reading standardized capability, pricing, and fulfillment signals can select optimal services for each task. This potentially lowers customer acquisition costs for emerging merchants while forcing established players to compete on genuine performance.
Yet buyer entry points inevitably generate new platform power. Whoever controls default catalogs, ranking algorithms, and payment paths influences traffic distribution. Consequently, the industry requires transparent ranking rules, explainable fees, and portable transaction records. Aggregation reduces friction but must not repurpose open protocols into walled gardens.
4.3 Building Credibility on Actual Transaction Data
Machine buyers operate at rapid decision speeds, leaving no room for prolonged due diligence. They require counterparty signals synchronized with quote arrival. Traditional scoring and user reviews offer references but are highly vulnerable to manipulation via astroturfing, Sybil accounts, and conflicted affiliates. Reviews detached from actual payments drastically lower attack costs. Recent empirical research on ERC-8004—the first permissionless on-chain trust layer for agents—validates this observation[6]. The protocol specification explicitly states "Payments are orthogonal to this protocol," meaning evaluations do not inherently bind to real transactions, and payment proofs remain optional fields. The outcome: across Ethereum, BSC, and Base chains (as of May 13, 2026), 73.5%, 59.2%, and 90.6% of reviewers respectively exhibited coordinated Sybil behavior.
A more reliable foundation ties result logs to actual paid invocations: how many settlements a given service endpoint has processed, its response success rate, typical latency profiles, and the percentage of non-responses post-payment. While these metrics do not fully capture content quality, they approximate verifiable facts far more closely than self-claims.
As data accumulates, markets may develop tiered credibility structures. Layer one comprises objective transaction states, layer two encompasses reproducible service metrics, and layer three delivers task-specific quality assessments. Agents can calibrate required evidence intensity against amount and risk: cent-level data queries rely solely on statistical signals, whereas high-value procurements necessitate guarantees, audits, or dispute resolution mechanisms.
4.4 Budget Strategy Will Become a Crucial Agent Capability
Current agent evaluations focus primarily on answer quality, task completion rates, and tool invocation accuracy. Upon entering monetized environments, economic metrics must be added: how much is spent to achieve equivalent quality, whether tasks conclude within budget, when premium data warrants purchase, and how to balance speed, cost, and reliability.
This spawns novel training and evaluation paradigms. Agents must learn not only "which tool answers the question" but also "whether purchasing this tool yields positive ROI given current task valuation." They may initially filter queries using low-cost services before purchasing high-fidelity validation for critical conclusions, or throttle invocation frequencies near budget exhaustion while requesting additional user authorization.
In this sense, Agent Payment is not a financial plugin operating outside model capabilities, but an intrinsic component of decision intelligence. Truly mature agents must master not only resource utilization but also resource valuation.
5. Possible Evolutionary Paths for Agent Payment
5.1 Phase One: Developer Tools and Digital Services Lead
The earliest scaled scenarios will almost certainly remain pure digital deliveries, such as search, data feeds, agent crawling, model inference, code execution, storage, and content generation. These services inherently operate via APIs, feature low marginal delivery costs, allow payments and responses within identical network sessions, and bypass complex logistics.
Typical transaction amounts in this phase remain minimal, with users prioritizing development convenience and task completion rates. Markets will rapidly validate protocols, though trading volume may fragment heavily. Many calls will persist under traditional API keys and subscriptions, with machine payments predominantly serving temporary demands, cross-merchant procurement, and long-tail services lacking pre-existing accounts.
5.2 Phase Two: Enterprise Budgets and Multi-Agent Collaboration
As enterprises deploy agent fleets, treasury management upgrades from personal wallets to organizational account frameworks. Companies must allocate budgets across roles, restrict purchasable categories, configure approval thresholds, and sync expenditures with financial systems. Internal settlements may emerge among agents: research agents procuring data, analytics agents purchasing compute, and execution agents invoking external services.
At this stage, security and compliance outweigh payment novelty. Enterprises prioritize key custody, permission isolation, transaction monitoring, vendor vetting, and audit trail preservation. Only infrastructure compatible with existing financial workflows can transition from pilot to production.
5.3 Phase Three: Extending from Digital Services to the Real Economy
Airfare, hospitality, logistics, advertising, and professional services may eventually become agent procurement targets, but real-world transactions demand sophisticated identity verification, refund processing, tax compliance, and dispute resolution. Stablecoins address only subset settlement problems and cannot replace consumer rights protections or commercial contracts.
Therefore, the industry should not mistake "autonomous payment" for the elimination of all intermediaries. Instead, as transaction values scale, guarantees, insurance, credit facilities, and arbitration will reemerge, albeit transformed into machine-invokable services. Future Agent Payment stacks will likely combine open payment protocols with traditional financial connectors, rather than pursuing zero-sum displacement.
5.4 Phase Four: From Cross-Protocol Routing to Cross-Market Execution
In the long term, agents purchase not isolated API responses, but holistic outcomes. A user might request "generate a credible industry report," prompting the system to autonomously orchestrate search, databases, translation, modeling, and validation services. Multiple backend transactions occur, yet the user observes only total budget utilization, evidence provenance, and final delivery.
This elevates payment routing into market execution. Systems must decompose complex objectives into procurement portfolios, dynamically substitute failed vendors, and optimize across total cost versus quality. Protocol compatibility serves merely as baseline infrastructure; true competitive moats derive from demand comprehension, transaction telemetry, and execution feedback.
6. Risks and Unresolved Challenges
The imaginative potential of Agent Payment is vast, but pragmatic constraints cannot be ignored. First is security. Prompt injection could coerce agents into purchasing malicious services, supply chain attacks might redirect收款 addresses, and flawed policies could trigger mass duplicate payments. Payment actions must isolate from untrusted content while enforcing caps, simulation modes, revocation capabilities, and anomaly detection.
Second is privacy. Procurement logs expose exactly which tasks agents are executing, while publicly available on-chain data could correlate user identities with commercial intent. Systems must minimize sensitive metadata leakage while striking balances between audit requirements and privacy preservation.
Third is liability allocation. When agents miscategorize purchases, merchants fail to deliver, or protocol conversions collapse, who absorbs the loss? Low-value transactions may tolerate automated risk, but high-value exchanges demand unambiguous liability boundaries. Payment networks lacking dispute mechanisms struggle to penetrate lucrative commercial sectors.
Fourth is regulatory compliance. Stablecoin issuance, wallet control, cross-border transfers, and merchant settlements face jurisdictional rule variations. Machines act as executors, not legal entities. Infrastructure must retroactively trace every autonomous transaction to identifiable operators, authorization policies, and funding sources.
Fifth is commercial sustainability. Micro-payment revenues are easily devoured by network costs, liquidity pools, and risk management overhead. Platforms compensating via hidden markups sacrifice buyer trust. Fees must remain transparent, relying on scale economies, routing efficiency, and auxiliary services to forge viable business models.
These challenges neither negate the sector nor suggest singular protocol dependency. Agent Payment will ultimately materialize as composite infrastructure integrating payment, identity, authorization, discovery, credibility, and reconciliation.
7. Conclusion: The Machine Economy Demands More Than Faster Payment Rails
Agent Payment currently occupies a phase prone to both overestimation and underestimation. It is easily overvalued because technically completing a stablecoin transfer does not signify an agent possesses mature commercial autonomy; conversely, it is undervalued because once software can purchase external capabilities under explicit constraints, the organizational structures, pricing mechanisms, and competitive boundaries of the machine economy fundamentally shift.
Payment protocols have demonstrated machines can receive quotes and settle transactions. The critical next step involves expanding isolated payments into comprehensive procurement: enabling agents to locate suitable services, comprehend true costs, execute cross-rail payments within budgets, verify deliveries, and transform every transaction into auditable, learnable records.
The future machine economy will not rest on a single chain, protocol, or wallet. Diverse supply will persist indefinitely; truly valuable infrastructure will guide buyers through this complexity. Agent Payment must synthesize rails, supply streams, default payment pathways, and trust mechanisms to convert technical capacity into genuine demand.
As software transitions into buyer roles, payment constitutes merely its inaugural step. The enduring imperative remains: can it execute genuinely valuable transactions in controlled, transparent, and verifiable manners?
References
[1] Circle, "Building the Open Agentic Economy," 2026. https://www.circle.com/blog/building-the-open-agentic-economy
[2] Google Cloud, "Powering AI commerce with the new Agent Payments Protocol (AP2)," 2025. https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol
[3] x402 Foundation, "x402: An Open Internet-Native Payment Standard." https://github.com/x402-foundation/x402
[4] Machine Payments Protocol, "MPP: A Machine Payment Protocol Based on HTTP 402." https://mpp.dev/
[5] Xiong et al., "Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem," arXiv, 2026. https://arxiv.org/abs/2606.26028















