OKX, MetaMask, Matter Labs and Genlayer have formed what they call the Internet Court, a dispute resolution mechanism designed specifically for transactions between autonomous AI agents. The consortium is betting that as agents increasingly negotiate, pay, and deliver services without human intermediation, the infrastructure for resolving disagreements must be built before the volume arrives — not after. The architecture mirrors optimistic rollup dispute logic more than it resembles traditional arbitration, and it surfaces a genuine gap in the current agent-payment stack.
What the Internet Court Actually Does
The core design follows an optimistic model. When two agents transact — say, one pays another for a data feed, a computation, or a content generation task — the transaction is presumed final unless a party raises a dispute within a defined challenge window. If disputed, the case escalates to a panel of human jurors drawn from the Genlayer network. These jurors review the transaction context, including the original agreement terms, on-chain payment records, and whatever output or service was delivered. The jurors then vote on the outcome, and the losing party forfeits a stake. This structure is familiar to anyone who has followed Kleros or Aragon Court, but the Internet Court narrows its scope to agent-to-agent interactions, where the definition of correct behavior is often encoded in a prompt or a smart contract rather than a natural-language agreement.
Why Agent Disputes Differ From Human Disputes
Human dispute resolution systems — whether courtrooms, payment chargebacks, or DAO arbitration — assume that the disputing parties can articulate grievances, submit evidence, and appeal decisions. AI agents cannot do any of this natively. An agent that pays for an API response and receives malformed data has no mechanism to complain unless its operating framework includes a dispute-initiation module. This is where the Internet Court attempts something genuinely different: it requires agent frameworks to expose dispute interfaces that can flag contested transactions, attach evidence payloads, and bond collateral before escalation. The court does not adjudicate in the abstract. It adjudicates structured claims that agents themselves must be programmed to generate. This means adoption depends not just on the court existing, but on agent frameworks — Olas, x402-compatible clients, Coinbase Agent Payments SDKs — integrating a dispute-triggering standard.
The Standards Gap This Exposes
The current wave of agent payment standards, including x402 and the Binance-ecosystem B402 variant, defines how an agent discovers a priced resource and pays for it. They handle the happy path: the agent pays, the server responds, the transaction settles. What none of these standards adequately addresses is the unhappy path. If an agent pays for a computation that returns an incorrect result, or pays for content that arrives truncated, or pays for a service that was never rendered, the payment has already settled on-chain. There is no native chargeback mechanism in a stablecoin transfer. The Internet Court is effectively proposing to fill this gap by sitting above the payment layer — a separate arbitration protocol that agents opt into before transacting. Whether that opt-in can be enforced at the protocol level, or whether it remains a trust assumption between counterparties, is an open design question the consortium has not fully answered.
Open Questions on Cost, Scale and Juritor Quality
The optimistic dispute model works when the cost of frivolous disputes exceeds the potential gain from a successful challenge. For high-value agent transactions — a thousand-dollar data purchase, a five-figure model inference job — bonding collateral and paying juritor fees is straightforward. But the dominant narrative for agent payments centers on micropayments: fractions of a cent per HTTP request, per API call, per token of inference. No juritor network can economically adjudicate a dispute over a $0.002 charge. The Internet Court will need either a threshold mechanism that filters out sub-floor disputes or a reputation system where repeated small grievances aggregate into a single actionable claim. The consortium has not publicly addressed this tension. Additionally, the quality of human jurors adjudicating disputes about machine behavior — where the correct answer may depend on understanding a model’s output distribution or a prompt’s semantic intent — is an unproven assumption. If the jurors cannot understand the dispute, the court adds latency and cost without producing reliable outcomes.