All rise: The Internet Court for AI Agents is now in session

In the near future, so-called AI agents will become tools accessible to everyone.

These agents run on the cloud, stand by on people's mobile devices, and assist them in handling various tasks, ranging from replying to emails, booking flights, to executing tax-loss harvesting for investment portfolios.

Currently, Robinhood users can already leverage AI agents to analyze sharp market fluctuations and automatically execute trades based on custom instructions. SAP's intelligent assistant Joule can help enterprise customers analyze inventory status, screen high-quality suppliers, and complete procurement processes. High-speed shopping agents like Amazon's Buy for Me can search the entire internet for the best-priced products, negotiate transaction terms with seller agents, confirm delivery time slots, and finalize orders.

01

A host of well-known AI and crypto enterprises including Anthropic, OpenAI, Coinbase, and Circle are racing to realize an inclusive future dominated by agents.

But new problems have emerged: if the sofa your agent ordered arrives in the wrong color, is delivered two weeks late, or is damaged, and the seller insists the damage occurred after delivery, how should you handle it?

Hidden within the grand blueprint of agent commerce is exactly this type of dilemma: it may incur high costs and could be difficult to avoid. While software can already replace individuals and enterprises to complete commercial activities such as procurement, bargaining, hiring, and payment, AI suffers from hallucinations, and business is never just a simple flow of funds—various unexpected errors are inevitable.

"Agent commerce is reaching a critical inflection point, but we are not yet prepared to address all the derivative risks," stated David Riudor, CEO and Co-Founder of the GenLayer Foundation. Headquartered in the Cayman Islands, the organization supports the operation of the new blockchain GenLayer and its core application Internet Court, which is designed to adjudicate various agent disputes. The entire platform operates without human intervention and is backed by 26 other crypto and AI enterprises, including industry giants such as crypto exchange OKX, wallet service provider MetaMask, and Binance's BNB Chain.

Despite its highly forward-looking and futuristic concept, Internet Court does not aim to completely replace human judges with various automated robots. It can be understood as an auxiliary system: helping agents draft contracts with clear terms, and if all parties cannot reach a consensus on the dispute outcome, an AI jury will verify the evidence and deliver a ruling within minutes.

Riudor noted that this technology is most applicable to scenarios involving small transactions. Hiring lawyers to safeguard rights in these transactions is simply uneconomical, yet ignoring them will still lead to losses. Albert Castellana, Co-Founder and CEO of GenLayer Labs—the team developing this blockchain—said: "We have no intention of competing with the existing judicial system. If the disputed amount is $10,000, hiring a lawyer to protect rights is not economically viable, and we aim to provide an alternative solution. Users can reach a resolution through this system for just a few cents."

The potential of this market could be enormous. According to data from Adobe Analytics, since October 2024, retail website visits driven by AI-referred traffic have increased by more than 14 times. McKinsey predicts that by 2030, the scale of global consumer transactions facilitated by AI agents could reach $3 trillion to $5 trillion. However, most of the emerging infrastructure supporting this nascent agent commerce currently only covers the ideal scenario where transactions are successfully completed: agents finish product selection and payment based on user needs, and the process ends after obtaining the corresponding goods or services.

02

At this stage, the application scenarios of Internet Court remain relatively limited.

The social platform Collective Memory encourages users to capture real-time photos, videos, and on-the-ground reports. The platform uses GenLayer to assess whether disputed visual materials are forged, such as a video recording a school under attack in Gaza or the situation on the streets of Tehran. Internet Court then reviews existing evidence related to the uploaded content, including upload time, geographical location, related submissions, and the user's past posting history, and issues a ruling on the authenticity of the material.

Ultimately, the designers of Internet Court hope that when disputes arise between AI agents, the system can automatically intervene to resolve them.

Imagine this scenario: at a small online clothing business, the owner delegates a large number of tedious daily tasks to several AI agents, which are respectively responsible for inventory management, ad procurement, and creative commission coordination. When the owner needs a new logo, her agent finds a designer who is also contacted through an agent. The two agents reach an agreement on the design proposal, price, and delivery date. After the logo is delivered, it appears fine at first glance, but reverse image search shows that the work is suspected of plagiarizing content from someone else's portfolio.

Internet Court is designed to provide solutions for such disputes: the two agents agree on all terms in advance and deposit funds into an escrow account; if any dispute arises, the case can be submitted to a jury for hearing before the funds are transferred.

The so-called jury mechanism is exactly where blockchain technology comes into play. The jury consists of five randomly selected blockchain participants (i.e., verification nodes), each running different AI models such as Claude, GPT, and Gemini. One of the five participants is randomly selected to act as the lead and propose a ruling opinion; the remaining members submit their votes without knowing the positions of others, then publicly state whether they agree with the opinion. If a consensus is reached, the system will open a 30-minute dispute window, during which agents or humans can challenge the result by posting a deposit. If the ruling is challenged, the jury will be expanded to 11 verification nodes, and will continue to scale up if necessary until a consensus is reached and no one raises objections to the ruling.

The design of this system is based on Condorcet's Jury Theorem. The theorem holds that under certain assumptions, the greater the number of independent reviewers, the higher the probability of reaching a correct conclusion. Nicolas de Condorcet, a philosopher and mathematician during the Enlightenment, proposed this theorem in 1785, and he eventually died in prison during the French Revolution. GenLayer believes that using multiple AI models can improve the system's anti-manipulation capability compared to relying on a single model or a single human arbitrator.

03

Although discussing disputes between agents sounds premature and somewhat abstract, Internet Court has already been deployed and is in the Beta testing phase.

Castellana revealed that the network can process approximately 350,000 on-chain transactions per day and complete 20,000 to 25,000 rulings. The project is scheduled to officially launch later this year and introduce a token to attract more verification nodes—a role open to everyone.

Riudor, head of the GenLayer Foundation, said that this ruling system can eventually be applied far beyond agent commerce, and can also be used in scenarios such as prediction markets. For example, the prediction market platform Polymarket relies on the UMA protocol, which escalates disputed results for voting by UMA token holders; Riudor stated that AI-assisted adjudication will achieve faster processing speeds.

Castellana said: "We have held discussions with several top prediction market platforms. These platforms are still waiting for our full launch, and are also evaluating us."

Andrew Hall, a professor at Stanford Graduate School of Business and research advisor for Andreessen Horowitz's crypto team, wrote earlier this year that using large language models as dispute adjudicators can help prediction markets achieve large-scale development, because these models are not at risk of being bribed and their performance iterations continue to accelerate. However, he also warned that large language models are prone to hallucinations and can be manipulated by carefully designed prompts or contaminated training data.

Lindsay Lin, current Chief Operating Officer and former General Counsel of New York-based crypto venture capital firm Dragonfly, also noticed the same contradiction: "Many large language models may face the problem of homogeneity, because they share training data and have common failure modes. In contrast, human judgments are often more independent."

However, Lin added: "People will tend to use AI to adjudicate disputes, especially small-value ones. Compared with human jurors, AI has lower costs and faster processing speeds; as the scale of agent commerce expands, a large number of disputes may arise. Therefore, it is necessary to establish a standardized protocol for agents, so that they can clarify collaboration terms and understand what remedies they can take when transactions fail to perform normally."

Other institutions have reached similar conclusions. In late June, the American Arbitration Association's International Centre for Dispute Resolution, the world's largest arbitration institution, released relevant standards for agents, namely the Legal Context Protocol. The institution maintains this standard together with Denver-based blockchain enterprise Integra Ledger, with Google, IBM, and a host of top crypto enterprises including Circle and Ava Labs participating to promote the introduction of this standard.

Of course, whether such standards can be implemented and promoted ultimately depends on two points: first, whether they can achieve widespread adoption, and second, whether AI models can be sufficiently reliable to alleviate people's concerns about hallucinations and bias.

However, infrastructure that allows agents to search for each other, hire each other, and complete payments has already begun to emerge. In recent weeks, OKX, a partner of GenLayer, and the NEAR blockchain team focused on the AI ecosystem have successively launched marketplaces where agents can hire other agents to perform various paid tasks, ranging from dataset acquisition to assisting with code reviews.

At the same time, physical courts have begun hearing disputes caused by violations of AI agents.

As one of the most high-profile AI-related litigations, in November 2025, Amazon sued AI enterprise Perplexity, alleging that its AI browser Comet logged into user accounts, disguised itself as a standard Google Chrome browser, and engaged in unauthorized transactions in violation of Amazon's Terms of Service. In March 2026, a federal judge in California issued a preliminary injunction prohibiting the Comet browser from shopping on Amazon; however, the appellate court later stayed the injunction to hear the appeal filed by Perplexity.

Regardless of the final ruling of the court, this case reflects a more prominent potential challenge in the field of agent commerce: without a general enforcement mechanism in place, how can millions of AI agents representing users to carry out various activities across platforms be regulated?

This article is from the WeChat Official Account "Forbes" (ID: forbes_china), author: Forbes, published with authorization from 36Kr.