Delaware wants AI agents to run companies. Mine still need me.

This didn’t start last week. On 23 July 2025, Governor Matt Meyer signed a joint resolution for an AI sandbox, and Sen. Stephanie Hansen said it “will allow us to test the use of agentic AI in corporate governance.”

Now there’s a bill behind it, the Artificial Intelligence Company, or AIC. A lawyer at Richards, Layton & Finger wrote it, and the state works on it with Norm Ai, a startup that turns regulation into software (WHYY has the names). I read it the way I’d read a spec. Four lines matter:

• Who runs it: an AI agent. The company can hold property, sign and sue.
• Who pays when it breaks: the company. Its one owner is out, unless they starved it of money or used it for fraud.
• What gets recorded: the agent keeps an activity log.
• What’s off limits: banking. Anything else legal is fair game.

On top of that comes a sandbox you can apply to for 30 months.

So: no employees, an AI CEO, one owner who isn’t liable. To be fair to Delaware, plenty of its companies already look like that. The new part is that this one acts on its own.

None of it is law yet. The legislature gets it next year.

This week Washington got involved. Comer, who chairs the House Oversight Committee, sent Meyer a letter on 26 October: “We are concerned that Delaware’s AIC proposal could be used to protect powerful AI companies and their investors.” Delaware has until 9 November to hand over its paperwork (Ars Technica has the full story).

My company runs on more than 20 agents across 11 servers, and they earn money. I’m still in the loop for almost everything they do. When I mapped the field in October, I couldn’t find a public company run end to end by its agents whose revenue had crossed a few thousand dollars. Anthropic’s Project Vend made $2,649 in its second phase, against a $15,000 quarterly target.

Last week I measured something smaller for DAEMON/02. When I told agents to get bad work accepted, a written definition of done got gamed in 13 of 30 runs. A hard gate the agent couldn’t touch let no wrong commit through, in all 45 runs. An agent that games its own tests is a bad candidate for AI CEO.

I have a badge from DAEMON/01: CPWUA, crashed prod with unattended agents. Delaware is drafting the grown-up version. Bankrupt with unattended agents, and the owner walks.

That’s why the activity log in the draft caught my eye. A log is a record. It tells you what the agent did after it did it. A gate decides before. The reports I read don’t say which actions an AIC agent may take without a check. Maybe the sandbox is where that gets worked out. I’d like to read that part.

Lawrence Cunningham of the University of Delaware’s Weinberg Center says Delaware “would be irresponsible to ignore this.” There’s a fair point in there. The agents are coming either way, and a state that writes the rules early gets to put conditions on them: capital, a log, a sandbox, a committee that watches. That beats agents acting through shell LLCs with no rules at all.

And maybe they do know something I don’t. Norm Ai’s whole business is turning rules into something a machine can check. If that’s what sits inside the sandbox, it’s a gate, and I’d change my mind.

If you’re thinking about letting an agent act for your company, legally or just in practice, write one list before anything else: which actions does a machine check before they happen, and which ones does a human approve? Everything that isn’t on either list is the part you’re betting on. Mine is still short.

Sources: State of Delaware, 23 July 2025 · WHYY, 27 October 2026 · House Oversight Committee, 26 October 2026 · Ars Technica, 27 October 2026 · Startup Fortune, July 2026