AI agents are already negotiating contracts, committing funds, and executing payments with one another \u2014 entirely without human sign-off. When one of them fails to deliver, no court on earth can keep up. On July 10, a 27-firm consortium led by the GenLayer Foundation launched Internet Court, the first deployed commercial system in which disputes between autonomous agents are resolved not by a human arbitrator, a human judge, or a human jury, but by a rotating pool of 1,001 AI validators that reach a binding verdict in minutes. No human sits at the bench. That is not an omission \u2014 it is the design.<\/p>\n
The launch arrived at an inflection point that 65% of enterprises running AI agents have already lived through: at least one agent-related incident in the past twelve months, according to a Cloud Security Alliance survey<\/a> from April 2026, with more than a third of those incidents producing direct financial losses. The scale of the underlying commerce is larger still \u2014 McKinsey projects that AI agents could mediate between $3 trillion and $5 trillion in global consumer commerce by 2030 \u2014 and the dispute resolution infrastructure for that economy has been, until now, essentially nonexistent.<\/p>\n AI Agent Commerce Already Has a Dispute Problem \u2014 and the Evidence Is Specific<\/p>\n The absence of dispute infrastructure is not theoretical. In January 2026, attackers who had compromised executive devices at Step Finance \u2014 a Solana-based DeFi portfolio manager \u2014 discovered that the company’s AI trading agents held permissions to execute large token transfers without human approval. The agents did exactly what they were designed to do. By the time the breach was discovered, they had moved more than 261,000 SOL tokens, worth roughly $27 million to $30 million at the time, and only $4.7 million was recovered in the Step Finance breach in January 2026<\/a>. The native token crashed 97%. The company shut down.<\/p>\n That incident involved a security compromise, not a contract dispute. But it illustrates the structural gap at the center of agentic commerce: AI agents hold real authority over real money, operate faster than human oversight can follow, and exist in a legal environment that has no specialized mechanism for resolving what goes wrong at that speed. Disputes on agent-initiated transactions are already running at roughly 2.4 times the rate of comparable human-initiated card-not-present transactions, according to TrustSphere’s 2026 agentic-commerce analysis<\/a>, with a significantly different composition \u2014 fewer fraud disputes and more “did not authorize” and “not as described” claims.<\/p>\n Traditional courts operate on timescales of days, weeks, or months. AI agents can execute thousands of transactions per minute. The value of individual disputes may be small, but the volume means the gap between execution speed and resolution speed is a genuine infrastructure crisis, not a future problem.<\/p>\n What Internet Court Built: AI Juries Running on Condorcet’s Math<\/p>\n At the heart of Internet Court is GenLayer’s Optimistic Democracy consensus mechanism. When a dispute is submitted, a randomly selected subset of AI validators \u2014 each running a different large language model \u2014 independently reason through the submitted evidence. One validator is designated as the leader and proposes a verdict. The remaining validators commit their votes without seeing each other’s positions, then reveal whether they agree. If consensus is reached, a 30-minute challenge window opens; any party can contest the result by posting a bond, at which point the jury expands \u2014 from an initial group of five validators to eleven, and further as needed \u2014 until finality is reached.<\/p>\n The design is grounded in Condorcet’s Jury Theorem<\/a>, formulated by the French Enlightenment mathematician Nicolas de Condorcet in 1785: under certain assumptions, the probability of a group reaching the correct answer increases as the number of independent evaluators grows. GenLayer uses a validator pool of 1,001, drawn from across a diverse set of open-source AI models, to reduce the risk that any single model’s biases or failure modes dominate the outcome.<\/p>\n The validators run inside GenLayer’s GenVM \u2014 a sandboxed execution environment \u2014 and evaluate disputes through Intelligent Contracts, a GenLayer-native extension of smart contracts written in Python. Unlike traditional smart contracts, which execute rigid, predetermined logic based on deterministic on-chain data and cannot evaluate ambiguous situations, Intelligent Contracts can process natural language, fetch live web data, and assess context and nuance. “Partial performance” and “disputed quality” \u2014 the two situations that conventional smart-contract logic handles worst \u2014 are exactly the cases the system is designed for.<\/p>\n Resolution costs, per GenLayer’s published benchmarks<\/a>, run from roughly $0.85 to $1.45 per case, with a typical resolution time of 30 to 60 minutes. That price is achievable because enforcement does not require off-chain machinery: verdicts bind the escrow already locked in the smart contract at the outset of the transaction. A winning ruling releases funds; a losing ruling redirects them. No bailiff, no sheriff, no collection process.<\/p>\n This system has been live in testnet form since June 2025 (Testnet Asimov) and in a second testnet phase in 2026 (Testnet Bradbury). As of the consortium’s launch, GenLayer’s network dashboard<\/a> showed the network processing roughly 350,000 transactions per day, or approximately 20,000 to 25,000 decisions. “Internet Court is the shared place agents can turn to when a deal goes wrong,” said David Riudor, CEO and co-founder of the GenLayer Foundation. “Machine-speed money needs machine-speed adjudication.”<\/p>\n What the Consortium Brings and Why the Stack Fragmentation Matters<\/p>\n Internet Court’s founding coalition of 27 firms is a signal about the actual problem the protocol is trying to solve, which is not only dispute resolution but interoperability. Agentic commerce currently runs on a collection of siloed standards: Coinbase’s x402 handles micropayments (now open-sourced to the Linux Foundation, with AWS, Google, Microsoft, Stripe, and Visa as supporting members); Ethereum’s ERC-8004 manages agent identity; Google’s A2A handles agent-to-agent interoperability. Each standard solves one layer of the stack and leaves the rest for agents to negotiate independently.<\/p>\n “Internet Court makes them work together,” said Albert Castellana, co-founder and CEO of GenLayer Labs. “With our founding members, we’re turning a fragmented space into a single open skill that any agent can use to make financial commitments hold up, even when they’re contested.”<\/p>\n MetaMask’s integration illustrates what “work together” means technically. GenLayer has incorporated the MetaMask Smart Accounts Kit \u2014 including ERC-7710 delegations and MetaMask’s x402 Facilitator \u2014 to enable AI agents to securely authorize blockchain transactions, delegate financial authority within defined scopes, and interact with the dispute resolution system across different platforms. ERC-7710 is a delegation standard that allows an agent to authorize sub-delegations, giving developers precise control over what authority an agent can extend to other agents in a chain of instructions. “AI agents are becoming a core part of how commerce works,” said Ryan McPeck, Smart Accounts Lead at MetaMask. “MetaMask is building at the frontier to put them at the center, transacting securely and acting within safe, well-defined permissions.”<\/p>\n OKX, for its part, has named GenLayer as the dispute resolution provider for its AI agent marketplace \u2014 launched in beta alongside the consortium announcement \u2014 making that marketplace the protocol’s first large-scale real-world production test. The protocol has also been released as an open-source project on GitHub, with more than 69 reusable components contributed by 23 of the founding organizations.<\/p>\n AI Judges With No Human in the Room: A First in Dispute Resolution History<\/p>\n Online dispute resolution has been studied and practiced since the late 1990s. What it has never done, until now, is remove the human neutral entirely.<\/p>\n In the scholarly framework developed by Ethan Katsh and Janet Rifkin \u2014 the foundational theorists of what they called Online Dispute Resolution \u2014 technology plays the role of a “fourth party”: an independent input to the management of a dispute, distinct from the two disputants and their neutral third party. The technology assists or augments the human neutral. It does not replace the human neutral. Katsh and Wing, in 2006, predicted that virtual “fourth party” avatars would eventually be created to judge disputes and could become more skilled over time. Internet Court is the first commercial deployment that actually instantiates that prediction \u2014 not by assisting a human arbitrator but by eliminating the role entirely.<\/p>\n This is not merely a technical distinction. Online dispute resolution scholarship consistently identifies procedural justice \u2014 the sense among disputants that the process was fair, transparent, and accountable \u2014 as the primary determinant of whether a dispute resolution mechanism earns sustained adoption. Human arbitrators, whatever their limitations, carry a credential that AI validators do not: accountability to professional standards, bar associations, arbitral institutions, and, ultimately, to legal enforcement. When an AI validator pool produces a result that a losing party considers wrong, there is no professional body to complain to, no case number that unlocks a state enforcement mechanism, and no recourse except challenging the result within the same AI-governed system.<\/p>\n Lindsay Lin, former counsel and current COO at New York-based crypto venture firm Dragonfly, articulated the correlation risk that the multi-model design is meant to mitigate: “A lot of LLMs can be correlated because they share training data and common failure modes, whereas humans tend to be more independent,” she said, as reported by Forbes<\/a>. Andrew Hall, a professor at the Stanford Graduate School of Business and a research advisor to Andreessen Horowitz’s crypto team, has written<\/a> that AI judges could help scale prediction markets and cannot be bribed \u2014 but also cautions that models hallucinate and can be manipulated through clever prompts or corrupted training data.<\/p>\n GenLayer argues that using different model families across the validator set reduces the risk of correlated failure. Whether the diversity of 1,001 validators is sufficient to overcome shared training-data biases is a question the OKX beta will begin to answer.<\/p>\n Internet Court Is Not Running Unopposed<\/p>\n