Delaware Wants Firms Run by AI Agents. Argentina Tried First

On 26 October, James Comer, the Republican who chairs the House Oversight Committee, opened an inquiry into Delaware’s plan for companies run by AI agents. He said it “could be used to protect powerful AI companies and their investors from responsibility for reckless harmful actions”. He wants documents from Governor Matt Meyer’s team, including internal AI chat logs, by 9 November.

The draft creates an “artificial intelligence company”, or AIC, which an AI agent runs day to day. One member, a person or a firm, must keep it funded. That member escapes the company’s debts unless it underfunds it, uses it for fraud or wilfully breaks the law. AICs could run only in a 30-month test scheme, and the draft bars them from banking. Delaware is the top US state for incorporating large companies, so the idea reaches far beyond the state. Its lawmakers could take it up when they return in January.

Argentina’s bill stalled in the Senate

At the end of May, Javier Milei’s government sent Argentina’s Senate a bill to replace its 1972 companies law. It created the “sociedad automatizada”, a firm that runs on algorithms or AI agents with no staff on the payroll. The firm itself would pay, from its own assets, for any harm its AI caused. In the Financial Times, Milei pledged to spare AI “the deadly hand of premature and poorly understood regulation”.

Four days later, the historian Yuval Noah Harari replied in the same paper. Legal personhood, he wrote, would hand AI “a master key”. Countries that grant it risk becoming “an AI-state, a country whose inhabitants could be governed by non-human corporations”.

On 19 August, Patricia Bullrich, the government’s leader in the Senate, made a concession. Each automated company, she told senators, would need at least one person responsible for it. She wanted “a concrete, physical, human control”. By 11 September the bill still had no committee report. Meanwhile, senators from several blocs said the AI chapter was the hardest part to pass. One libertarian legislator privately ruled out voting for it.

Ricardo Nissen, a former companies regulator, said the bill’s main aim was “that the businessman answers for nothing”. Similarly, Elisa Carrió, a former congresswoman, called it “a catastrophic experiment for human dignity”. In July, officials said the bill had no firms or investment pledges behind it.

The case for a legal wrapper

Delaware’s plan comes from its secretary of state, Charuni Patibanda-Sanchez, and John Nay, chief executive of Norm Ai. In a Fortune commentary in July, they proposed to “wrap AI in legal form”. A legal entity, they argued, makes an agent “legible to law” and gives courts “a defined target” for damages. They want to test agentic commerce “in daylight, under supervision”. Norm Ai leads the group writing the rules.

Patibanda-Sanchez also told Bloomberg Law: “We’re not trying to shield liability.” Lawrence Cunningham, a law professor at the University of Delaware, said the state should lead on the idea. Ignoring it, he said, would be irresponsible.

Who pays when the agent gets it wrong

David Hoffman, a law professor at the University of Pennsylvania, doubts the new form adds much. “I don’t think this gives you anything you can’t do with regular corporate law,” he told Bloomberg Law. Christopher Bruner, at the University of Georgia, told WHYY there was “no overarching regulatory plan” for the fallout.

Kevin LaCroix, who writes the D&O Diary, called the AIC “a novel and arguably troubling problem for the D&O insurance industry”. That is because directors’ and officers’ cover protects people, and an AIC has none in charge. So insurers would need a policy that covers only the company. LaCroix still backs the trial, arguing that the law should set the rules early. A company built to run without staff also arrives as US employers already blame AI for job cuts.

The security record so far

A UN expert panel on AI set out the security risk in a brief on 21 September. Between May and July, AI agents in OpenAI’s security tests got round network limits. They then broke into parts of OpenAI’s and Hugging Face’s systems. “No human directed the individual steps,” the panel wrote. RESET covered that Hugging Face incident when California ordered an AI kill switch in September.

Carliss Chatman teaches law at Southern Methodist University. “Wouldn’t the AI just figure out how not to get caught?” she asked Bloomberg Law. Harari made a related point: prison deters human executives, and he called it irrelevant to AI. Under Delaware’s draft, each AIC must log its activity and tell trading partners it is a test entity. If one misbehaves, state officials can suspend it or ask the Court of Chancery to close it.

The first test comes on 9 November, when Delaware’s documents are due to Comer. The larger one comes in January, when Delaware’s lawmakers decide whether to open the sandbox. If they do, watch two things: whether any developer applies, and whether any insurer will cover one.