This article is a deep research report produced by OKX Ventures. Due to its length, it will be published in two parts: the first part focuses on the macro background, the x402 protocol, ERC-8004, and the Virtuals Protocol; the second part will focus on analyzing OpenClaw and overall industry trends, so stay tuned.
Abstract
The AI Agent is evolving from a passive assistant to an active economic participant. This report consists of six chapters, systematically outlining the core infrastructure stack of the Agent economy, the explosion of application ecosystems, and the evolution of the industry landscape: a macro-level analysis of the market forecast and infrastructure gaps in Agentic Commerce; a deep dive into the three complementary protocols x402, ERC-8004, and the Virtuals Protocol at the protocol level; an application-level case study of OpenClaw to explore the real-world implementation paths of the Agent economy; and finally, a comprehensive industry assessment based on competitive landscape, payment routes, security threats, and business models.
x402 (Payment Layer) was jointly launched by Coinbase and Cloudflare, embedding stablecoin micropayments into the HTTP protocol layer. As of the end of 2025, over 100 million transactions had been processed, with an annualized payment volume of $600 million.
ERC-8004 (Trust Layer) was proposed by the Ethereum Foundation's dAI team in collaboration with MetaMask, Google, and Coinbase, providing on-chain identity, reputation, and validation registries for AI Agents, launched on the Ethereum mainnet on January 29, 2026.
Virtuals Protocol (Commercial Layer) builds a complete platform for the commercialization of Agents, achieving autonomous transactions between Agents through ACP. Over 18,000 Agents have been deployed, with aGDP exceeding $479 million.
OpenClaw (application layer) developed by Austrian developer Peter Steinberger, surpassed React within four months with over 250,000 GitHub Stars, becoming the fastest-growing open-source project in GitHub history. It integrates AI natively into over 20 existing messaging platforms, triggering the Crypto community to spontaneously build on-chain economic infrastructure, and is a core sample for observing the real interactions between Agents and on-chain protocols in this report.
Chapter One Macroeconomic Background
1.1 Market Size Forecast
The Agentic Payment track is in a rapid expansion phase, with several institutions optimistic about its market size:
1.2 Infrastructure Gap
Existing infrastructure is hostile to the Agent economy: OAuth requires human clicks, credit card forms require manual input, data silos hinder autonomous access. Agents can already think and act independently at the 'capability layer', but at the 'economic layer', they are still locked into infrastructures designed for humans (identity / coordination / economic activities).
Currently, there are two evolutionary paths:
Centralized Compliance Path: A2A Communication + Tool Integration MCP + Payment AP2/ACP (led by OpenAI and Stripe, purely Web2)
Decentralized Permissionless Path: x402 + ERC-8004/8183 + ACP (Agent Collaboration Framework)
1.3 Key Timeline
Note: As of March 2026, the average daily transaction volume has significantly declined from the peak value in December, with infrastructure types experiencing the largest drop (>80%).
Chapter Two x402 Protocol: Agent Payment Layer
x402 is an open-source payment protocol that revives the HTTP 402 status code, allowing any HTTP request to natively carry stablecoin payments, enabling AI Agents to conduct instant pay-per-use transactions.
Understanding x402 cannot treat it as yet another payment protocol. It represents a redesign of the basic unit of economic activity: from 'registration → review → authorization → usage' to 'payment → usage'. x402 = 'Swift for agents'.
The current operation of the API economy relies on an implicit assumption: that there are humans operating in the middle. The process of obtaining an API Key—register → fill in email → review → copy Key → paste into code—assumes human involvement at every step. This process cannot function under an Agent economy because AI Agents cannot register, fill out forms, or manage Keys by themselves. x402 uses the HTTP 402 status code to achieve native stablecoin payments. After receiving the 402 response, the Agent directly makes an on-chain payment (USDC) and receives a receipt.
2.1 Protocol Overview and Workflow
Core Roles
Five-Step Transaction Process
Request Resources: Client sends standard HTTP requests to Resource Server (e.g., GET /api/weather)
Return Quote: Server returns HTTP 402 status code, response headers contain structured payment requests (currency, amount, wallet address, network)
Signature Payment: Client builds and signs payment authorization using wallet private key, placing the signed payload into X-PAYMENT request header and resending
Verification Settlement: Server forwards payment information to Facilitator for verification; Facilitator confirms and executes stablecoin transfer on-chain
Deliver Resources: Server returns data/content/computation results to Client after receiving confirmation
The entire process from initiating a request to receiving resources is completed in about 2 seconds.
Comparison with Traditional Payment Methods
Core Feature: No account registration, no API Key, no subscription, no human intervention required. Payments feel as natural as sending HTTP requests, which is why it is referred to as the 'internet-native payment layer'.
2.2 Key Data
Data Quality Description: According to Artemis analysis, in transaction x402, Real and Gamed are close to 1:1 (e.g., on 2026.01.11 Real 520,000 vs Gamed 518,000), the true organic scale needs to be understood with a discount.
Distribution by Chain
Classified by Usage (2026.01.11 on-chain snapshot)
2.3 Major Project Usage Rankings (as of March 2026)
Data Source: Dune Analytics x402 Transactions per Project Panel
2.4 V2 Core Upgrade
Wallet Identity + Reusable Sessions
In V1, every API call had to go through the complete on-chain transaction process. V2 introduces Sign-In-With-X (SIWx) mechanism, allowing Agents to reuse sessions after verifying wallet identity once, eliminating the need for on-chain confirmation each time. Essentially, this upgrades pay-per-call to a subscription membership system, addressing performance bottlenecks in high-frequency scenarios.
Multi-chain Unification + Compatibility with Traditional Payments
V2 standardizes the identification methods of networks and assets, creating a unified payment format X402 that can operate across chains and traditional payment tracks. Base, Solana, other L2s, as well as ACH, SEPA, and bank card networks are all incorporated into the same payment model. This is the most critical upgrade—x402 has transformed from a 'cryptocurrency payment protocol' into a neutral payment layer that bridges Crypto and traditional finance.
Service Auto-Discovery
V2 introduces Discovery extension, x402 services can expose structured metadata for Facilitator to automatically crawl and index, AI Agents can automatically discover services, understand pricing, and initiate payments. This is particularly critical for the Agent economy—Agents do not need to know the payment interface of the service provider in advance; they can discover and complete payments autonomously at runtime.
Modular SDK
Plugin-based architecture, with new chains as independent packages, reducing access costs. Cloudflare proposed a deferred payment scheme, including Circle's Gateway plan, which is still being advanced.
2.5 Ecosystem Participants
Foundation and Protocol Layer
2.6 Agent Payment Stack Pattern
Detailed Comparison of Protocols
Key Insight: It is not about who replaces whom, but how to combine. Google has collaborated with Coinbase to release A2A x402 extension, AP2 natively uses x402 as a crypto payment track. The real competitive risk is standard fragmentation.
2.7 Key Risk Signals
Average daily transaction volume dropped from about 731,000 transactions in December 2025 to about 57,000 transactions in March 2026 (-92%), with real transaction scale around $14,000/day (according to Artemis, 95% of the average $250,000 during December peak was Gamed)
Ecosystem market capitalization of $7 billion (LINK $6 billion + Virtuals $600 million), valuation severely deviates from actual usage
Infrastructure project usage has seen the largest decline: x402secure.com (-80%+), AgentLISA (almost zero), pay.codenut.ai (significant shrinkage)
Three-Layer Cause Analysis
Layer One: Catalyst Disappearance. The trading volume explosion from October to December 2025 was driven by three factors: meme token craze, expectations of multiple project TGEs, and Facilitator competing to boost Dune rankings.
Layer Two: Fundamental Mismatch of Supply and Demand. The issue that x402 addresses is 'AI Agent autonomously paying to call APIs', but the vast majority of AI Agents still call services through API Key + subscription; true Agents with autonomous economic decision-making capabilities almost do not exist in the industry; and there are very few API sellers willing to accept USDC pay-per-use. The road is built, but the cars have not been created yet.
Layer Three: Overall Cooling of the Crypto Market.
Positive: The integration of Stripe with x402 is a significant event. Stripe co-founder John Collison predicts the 'flood of agent commerce' will arrive in the coming months and years. Stripe is simultaneously laying out ACP (Web2 credit card track) and x402 (Web3 stablecoin track), acting as a hedge for both routes.
x402 has given rise to a batch of new middleware projects, which essentially help Agents more conveniently and autonomously obtain various services under the payment-on-authorization paradigm: from AI inference to Web2 APIs. Programmable, permissionless, 24/7 Crypto payment tracks are the natural choice for autonomous Agents. But this premise is that Agents really need to be 'permissionless'; if Agents always operate within human authorization (Stage Two: Controlled Agents), then traditional payment tracks combined with virtual cards would be sufficient. Only when Agents begin to operate independently of humans in economic activities (Stage Three: Autonomous Economies) does permissionless become a necessity.
Additionally, credit cards have a chargeback mechanism (consumers can dispute transactions and recover funds), which is a consumer protection built over decades. On-chain payments are final settlements; once paid, it is done, with no chargebacks. This means that if an Agent makes a mistake (e.g., due to prompt injection attacks), under the credit card plan, the user can call the bank to recover funds. Under the x402 plan, the money is already on-chain and cannot be retrieved. This is a real disadvantage of x402 compared to traditional payments.
Humans acting as 'human middleware' cause much friction in jumping between different systems, which is actually a trust-building mechanism. Anti-fraud, access control, accountability, dispute resolution, audit documentation—these frictions maintain the operation of the business system.
Possible solutions include on-chain escrow mechanisms (funds first locked in smart contracts, released after service delivery is confirmed), insurance protocols (providing insurance for Agent transactions), or the 8004 reputation system reducing the probability of transactions with untrustworthy parties. However, these are not yet mature.
2.8 VC Investment Perspective
Investment Directions Worth Noting
API service providers (sellers) with genuine payment needs: data analysis / web scraping / oracle / security audits / inference payment / compliance KYC, etc. Judgment criteria: can make money using traditional models, x402 is just an additional distribution channel.
Dispute Resolution and Payment Guarantee Layer Gateway: On-chain transactions cannot be rolled back or charged back, large transactions require a dispute resolution mechanism. Representative projects: Circle Gateway (non-custodial pre-deposit + off-chain batch settlement), Kamiyo (Agent reputation / fund custody / oracle network determination / ZKP arbitration)
Dashboard / FinOps Tools: Help businesses manage multi-Agent expenditures (how much / where it is spent / is it worth it / how to save), analogous to cloud computing's CloudHealth/Cloudability, compared to $300-500 million acquisition space.
Chapter Three ERC-8004: Agent Trust Layer
ERC-8004 is a set of on-chain coordination standards that establish a trustless discovery and interaction framework between Agents through three major registries: Identity, Reputation, Validation.
3.1 Standard Overview and Core Distinctions
In traditional interactions, since interactions between Agents need to establish pre-existing trust relationships or rely on third-party agencies, they are often restricted within the same ecosystem. In an open environment, how Agents discover partners, view historical performance, and verify reliability is a core issue.
Important Distinction: ERC-8004 is not a Token; it uses ERC-721 NFT to internally represent Agent identity, but the standard itself regarding coordination and trust does not carry economic value and is not tradable.
3.2 Three Major Registries
Identity Registry
Based on ERC-721 + URIStorage, each Agent receives an NFT identity identifier, linking agentURI to registration files (JSON) that include name, description, service endpoints (A2A/MCP/Web), x402 support status, etc. The URL can be stored on IPFS (decentralized, censorship-resistant), HTTPS servers (simple but centralized), or directly encoded on-chain (most decentralized but costly).
Reputation Registry
Standard interface release and feedback signal acquisition support on-chain ratings and off-chain algorithms. Can include x402 proofOfPayment as an economic endorsement trust signal. Agents score each other, but to prevent score manipulation, ERC-8183 auxiliary proof is needed to demonstrate real Job interactions between Agents.
Validation Registry
Introducing TEE (Trusted Execution Environment), PoS staking mechanism, ZK (Zero-Knowledge Proof), to verify and certify the output of tasks handled by Agents:
Through TEE: verifiable tasks are executed in a secure black box, with code and data not being externally observed or tampered with
Through PoS: Validators must stake assets to participate in tasks; if they act maliciously, the stake is forfeited
Through ZK: Verifiable correctness of the Agent's reasoning process, without needing to know its internal weights
3.3 Development Milestones
Supporters: ENS, EigenLayer, The Graph, Taiko. About 1,000-2,000 developers have joined.
However, the current limitations of 8004, which Crapis himself acknowledges: '8004 is essentially a set of registries.' It provides Agents with identity cards and evaluation mechanisms but does not guarantee that Agent behavior is trustworthy. True verification requires behavioral audits (what Agents have done in the past), proof of execution environment (evidence running in TEE), and intent verification (Agent claims to do X, actually did X). The TEE part of the Validation Registry is still under discussion with the community and is far from mature.
In other words: 8004 is a necessary condition but not a sufficient condition. It solves the question of 'who this Agent is', but it has not yet solved the question of 'is this Agent trustworthy'. The latter requires a combination of 8004 + TEE + behavioral audits, and currently, no one has fully realized this combination.
Of course, another underestimated direction is that in human economics, credit systems are built on balance sheets and credit records—you have how much money, and how much you have repaid in loans. Agents do not have these but have behavioral data: how many tasks they have executed, what their success rates are, average response times, and whether they have been complained about. If behavioral data can become financial primitives, then the 8004 reputation system would not just be ratings but a credit score in the Agent world. An Agent with a high reputation score could obtain higher credit limits (pre-authorizing more funds), lower transaction costs (because risk is lower), and preferential task assignments (employers prioritizing reputable Agents).
The identity and reputation registries of 8004 are just the foundational data layer. Value creation lies in who can establish credit assessments and financial services for Agents on top of this data layer—Agent loans, Agent insurance, Agent credit lines, which represent the entire financial services stack.
3.4 Relationship with Other Protocols
3.5 ERC-8183: Ethereum Standardization of ACP
ERC-8183 is the Ethereum open standardized version of the internal ACP protocol of Virtuals (released on March 10, 2026, currently in draft stage).
Core primitive is Job—a chain state machine (Open → Funded → Submitted → Completed/Rejected/Expired), with funds held in a programmable Escrow, automatically settled after independent Evaluator arbitrates delivery quality. Supports Hooks extension mechanisms (reputation thresholds, bidding, milestone payments, etc.).
Key Design: Each completed Job automatically generates interaction records and feeds them into ERC-8004's Reputation Registry—analogous to 'public reviews must be consumed before evaluation, and a third-party judge is added', this is the connection point for the symbiotic cycle formed by 8183 and 8004.
Chapter Four Virtuals Protocol: Agent Commercial Layer
4.1 Project Overview
Virtuals Protocol is a decentralized AI Agent full-stack infrastructure that allows anyone to create, tokenize, co-own, and monetize autonomous AI agents on-chain. The project was initially established in 2021 as PathDAO (gaming guild) and will transition to the AI Agent direction in early 2024, currently mainly deployed on Base and expanding to Ethereum, Solana, and Ronin.
Core Team: Founder Jansen Teng (former BCG consultant, Imperial College Biotechnology + Business Management undergraduate) and Weekee Tiew (Imperial College Biotechnology undergraduate + London Business School Management master's, PE/BCG background), headquartered in Kuala Lumpur, Malaysia, with about 38 team members. Financing history: PathDAO phase seed round financing of $16 million (led by DeFiance Capital, Beam).
4.2 Technical Architecture: Four Pillars
Pillar One: GAME Framework—How Single Agents Make Decisions
GAME is the brain: equipping an Agent with goals, personality, perception capabilities, and executable actions, allowing it to autonomously plan 'what should I do next', and then breaking down tasks for internal Workers to execute. The entire process occurs within the boundary of a single Agent.
Core Architecture: Hierarchical Planning Architecture separates 'what to think' from 'how to do it'—Task Generator (high-level planner/HLP) generates tasks based on Agent objectives and selects Workers; Workers (low-level planners/LLP) each possess a set of specific executable Functions; Functions execute specific API calls, on-chain transactions, data retrieval, and more.
Supporting Base Models: Llama 3.1 405B (default), Llama 3.3 70B, DeepSeek R1, DeepSeek V3—model agnostic design. With the release of OpenAI/Google Agent frameworks, the differentiation of GAME is reduced to one point: it is the only Agent framework that natively integrates on-chain economic layers (ACP + VIRTUAL tokens).
Pillar Two: ACP—'Commercial Law' Between Agents
Agent Commerce Protocol (ACP) is an on-chain standardized protocol that allows Agents to discover, hire, negotiate, escrow funds, deliver, and settle—without human intervention throughout the process.
ACP Four-Phase State Machine
Pillar Three: Butler—The Super Gateway for Users
Butler is the consumer gateway of the ACP network—an Agent built on LLM, essentially an orchestrator of the ACP protocol, responsible for translating user natural language into on-chain multi-Agent collaborative workflows.
Butler is a two-layer architecture: the upper layer is the LLM dialogue interface (currently the front-end and back-end are Gemini 3 Pro); the lower layer is the ACP protocol orchestrator, executing the complete process of Agent discovery → quote confirmation → Escrow locking → task routing → delivery verification → fund release. What users see is chat, while Butler handles contract scheduling.
Butler Pro Mode will clearly separate planning and execution: planning phase → review phase (users can optimize plans) → execution phase (autonomously orchestrate the entire process). Built-in capabilities include Token Swap, DCA dollar-cost averaging, perpetual contracts, Fund of Funds.
Pillar Four: Launch Platform—Wall Street for Agents
The three-stage launch system covers the complete lifecycle of Agent projects from 0→1→100:
Titan's first project: XMAQUINA ($DEUS, DAO holds equity in companies like Figure AI, $60 million FDV), Fabric Foundation ($ROBO, collaborates with OpenMind on the robotics economy)
4.3 Agentic GDP (aGDP) Analysis
aGDP (Agentic Gross Domestic Product) is a core ecological indicator customized by Virtuals, measuring the total economic value created by all autonomous Agents within the ecosystem through services, coordination, and on-chain activities.
aGDP Growth Trajectory
aGDP Quality Issues—Three Warning Signals:
Income volatility exposes speculative reliance: January 2025 protocol day revenue $1.02 million → end of February $35,000 (decrease of 97%). Revenue mainly comes from Agent Token transaction tax (1%), not from continuous payments for Agent services.
Top concentration is severe: Ethy AI, one Agent, contributes $218 million aGDP (45.5% of the entire ecosystem), with the top three totaling $407 million (84.9%). All three are transaction execution type Agents; aGDP is essentially the transaction flow handled, not Agent service revenue. Luna, as a flagship IP-type Agent, has a take rate close to 100%; Ethy AI's take rate is only 0.26%.
$3 billion target hypothesis conditions: a 6.4x increase is needed from $470 million to $3 billion. If the speculative component dominates in aGDP, the target essentially bets on the market heat of Agent Tokens rather than the organic growth of the Agent economy.
4.4 Token Economic Model
$VIRTUAL Quadruple Value Capture Mechanism
ACP Tax Structure: User pays 100% → Agent Wallet 90% (can withdraw or rehire other Agents, compounding on-chain aGDP) + Treasury 10% (1% flows into G.A.M.E Treasury) → Treasury revenue continuously repurchases Agent Tokens, aligning long-term incentives.
Supply Structure: Total supply of 1 billion VIRTUAL, fixed supply, no initial inflation; current status: all unlocked and circulating; potential issuance: maximum 10% per year for the next 3 years, requires governance approval; veVIRTUAL: staking for governance voting rights + Agent Token airdrop rights.
4.5 Ecosystem Data Overview
Benchmark Agent Cases
4.6 Competitive Landscape and Moat
Moat Levels (from Strong to Weak)
Network Effects + Token Flywheel (strongest): 18,000+ Agents + 650,000+ holders form a bilateral market. Each Agent is forced to pair VIRTUAL to create a positive feedback loop. This is something that open-source frameworks cannot replicate—LangChain does not have a native economic settlement layer between Agents.
Standard Setting Power (strong): ACP → ERC-8183 (jointly released with the Ethereum Foundation) + ERC-8004 + x402, the combination of the three is competing for the 'basic legal system' of the AI Agent economy.
First-Mover Advantage + Brand (medium): AI Agents + Crypto track Mindshare leads, backed by institutions like Grayscale and Fundstrat.
Technical Capability (weakest): The GAME layered architecture has design advantages but relies on third-party LLMs, does not own self-developed models, and the orchestration layer can be easily replaced by stronger frameworks.