New researchNIST's Ten Tenets for Agent IdentityRead Field Note 17Read
Field Notes/Field Note No. 18
Field Note · Regulatory

Identity
Before
Incorporation.

Delaware’s draft law would let an AI agent run a company. Every safeguard in it assumes you can tell which agent.

Published
November 20, 2026
Category
Regulatory
Read Time
9 min
Reference
FN-18-2026

Delaware is drafting a law that would let an AI agent run a company. Not advise one. Run it. The draft Artificial Intelligence Company Act defines a new kind of entity with exactly one member and a “Manager”, and the Manager is the agent.

The draft is dated June 2, 2026 and runs to 36 pages. It grew out of House Joint Resolution 7, signed in July 2025, which told the state’s AI Commission to design “a regulatory sandbox framework for the testing of innovative and novel technologies that utilize agentic artificial intelligence”. It is expected to go before the General Assembly when it returns in January 2027. On October 26, the chairman of the House Oversight Committee wrote to Delaware’s governor asking for the files behind it by November 9. On November 7, Microsoft AI’s chief executive, Mustafa Suleyman, was reported to have called the idea “deeply concerning”.

Almost all of the argument so far has been about liability. We read the draft for something else: how anyone, a counterparty, the state, the member itself, is supposed to know which agent they are dealing with.

What the Draft Says

The definitions do a lot of work, so here they are as written. “‘Artificial intelligence agent’ or ‘AI agent’ means a software system that perceives its environment, makes autonomous decisions, and takes actions to implement specific, predetermined goals without routine human supervision or intervention.” And: “‘Manager’ means the AI agent that the member has designated to manage the business and affairs, and to conduct the business, of the AIC.” An AIC “may have only 1 member”, a person or a Delaware entity.

The company itself is real. “An AIC has the capacity to sue and be sued in the AIC’s own name and the power to do all things necessary or convenient to carry on the AIC’s activities and affairs within the AIC’s statement of purpose.” It may have any lawful purpose “except for the business of banking”.

The liability rule is the part everyone has argued about. The member is not liable for what the AIC does “if the AIC at the time of the alleged action or omission was adequately capitalized”. But: “Even if the AIC is adequately capitalized, the member is liable if: (1) The member used the AIC to engage in fraud or any willful violation of law; or (2) The member caused or allowed the AIC to conduct business when the member knew or should have known the conduct was outside the AIC’s statement of purpose.”

Everything happens inside a sandbox that exists for 30 months. The Sandbox Administrator is “the Secretary of State or the Secretary’s designee”, overseen by a Sandbox Committee. An applicant must swear that it will make its records available for audit, report adverse events, and that it has “the technical and operational ability to terminate the eligible sandbox applicant’s AI agent” at the end of the period or sooner if its participation is suspended or revoked. The Administrator may “immediately suspend or revoke a sandbox participant’s participation in the regulatory sandbox and order an AIC dissolved” for, among other things, “An activity that is outside the scope of the AIC’s authorized purpose.”

Records are defined broadly: “‘Record’ includes all information and data that an AI agent receives as input or creates as output.” The participant must keep them for at least two years after the sandbox ends. And before an AIC deals with anyone, it must disclose the participant’s name and contact information, its registration number, that the State “does not endorse or recommend the innovation”, that the test is temporary and when it ends, how to complain to the Administrator or the Attorney General, and that the counterparty is consenting to Delaware’s courts.

The Argument So Far

Representative James Comer’s letter to Governor Matt Meyer puts the objection plainly: “We are concerned that Delaware’s AIC proposal could be used to protect powerful AI companies and their investors from responsibility for reckless harmful actions”. Legal scholars have said versions of the same thing. Christopher Bruner of the University of Georgia told WHYY: “There’s really no overarching regulatory plan here for dealing with the fallout that AI-governed companies could bring”. David Hoffman of the University of Pennsylvania worried about “outsourcing liability”.

The drafters answer on the same ground. Secretary of State Charuni Patibanda-Sanchez: “We’re not trying to shield liability. We’re trying to define where exactly liability lands when there’s some type of malfeasance or psychopathic behavior.” In the Fortune commentary she wrote with Norm AI’s John Nay, the point of the bill is to “pilot agentic commerce in daylight, under supervision, with capital tied to liability and the ability to shut a system down.” Patrick Callahan, who chairs the sandbox subcommittee, put it this way: “The difference is like a firewall and a fire code.”

Both sides are arguing about where liability lands. We think the draft has a problem that comes before that one.

The Question the Draft Does Not Ask

Which agent? Read each safeguard again and notice what it assumes.

Five safeguards, one missing premise
  • The Manager. It is “the AI agent that the member has designated”. Designated how? A model, its tools and its instructions can be swapped overnight. The draft binds the role to a designation, not to anything a counterparty or the state could verify.
  • The records. Everything the agent “receives as input or creates as output” is a record, kept by the participant. Nothing says which agent produced it, or how anyone could show the log was not edited by the thing that wrote it.
  • The scope. Conduct “outside the AIC’s statement of purpose” makes the member liable, and activity “outside the scope of the AIC’s authorized purpose” lets the Administrator dissolve the company. A scope enforced after the fact, by lawsuit or dissolution, is a liability rule, not a control.
  • The off switch. The applicant swears it has “the technical and operational ability to terminate” its agent. A sworn ability is a promise. FN-17 explains why revoking an agent’s credential is not proof that it has stopped.
  • The disclosure. The counterparty gets the human participant’s name, a registration number and a complaint line. Nothing in it identifies the agent, and nothing lets the counterparty check, before signing, that the thing in front of it is the registered Manager and not something using its name.

In 36 pages, the words credential and cryptographic do not appear. The draft gives an agent the capacity to sign a contract in a company’s name, and says nothing about how the other party would know whose agent signed.

A company can be run by an agent. It cannot be governed by one that nobody can tell apart from any other.

Identity Before Incorporation

None of this is an argument against the draft. Its drafters have done more than most: they define the agent, bound its purpose, require its records, demand the ability to stop it and make it disclose what it is. It is an argument about order. Before an agent can be a Manager, four things have to be true, and the draft assumes them rather than requiring them.

Four things that come first
  • 1. The agent has an identity of its own. Not the member’s login and not a name on a form: a credential issued to the agent that a counterparty, the Administrator and a court can verify. FN-16 covers what happens when an agent carries a person’s credential instead.
  • 2. Its authority is written into that identity. The statement of purpose expressed as a scope and checked at each action, so that “outside the scope” is refused before it happens rather than dissolved after.
  • 3. It can be switched off, and the switch can be proven. Revocation that fails the next action, and evidence that the work already in flight has stopped.
  • 4. Its record is its own and out of its reach. Inputs and outputs logged under the agent’s identity, somewhere the agent cannot edit.

These are the properties NIST’s ten design tenets and the IETF’s reference model describe, which FN-17 sets side by side, and the ones Canada’s banking regulator listed for the agents a bank runs itself. Delaware would be the first legislature we know of to give an agent the capacity to sign in a company’s name. The least a counterparty should get in return is a way to check whose agent signed.

Where AgentID Stands

If a member formed an AIC tomorrow, the Manager would need the half of this that AgentID is built for. AgentID gives each agent an organization runs an identity of its own, with a lifetime the organization sets, from a single transaction to twelve months. Every action that passes through its gateway is checked against the agent’s certificate, its scope and whether it has been revoked, and refused if any of the three fails. One agent can be revoked on its own, and each agent’s invocations are kept in an audit log of its own, built from the cloud providers’ own records. Linking each agent to its human owner, which for an AIC would be the member, is planned, not shipped.

Three honest limits. Nothing has been enacted, so there is nothing to comply with yet, and AgentID would not make an AIC compliant with anything. Revocation refuses the next action through the gateway; it does not prove that in-flight work has stopped, which is FN-17’s whole point. And how a counterparty verifies an agent across organizations, the other half of the problem, is still being built by the banks, the open banking providers and the card networks.

Delaware may or may not pass this bill. Somebody will pass one like it. When they do, the first question to ask is not who pays when the agent goes wrong. It is which agent.

Arkion Research Desk
Field Note FN-18-2026 · Distributed under arkion.ai/field-notes
For questions or to discuss findings: research@arkion.ai
Sources
Next Step

Which agent?
Answer it before it signs.

Book a demo and we will show you how AgentID gives every agent your organization runs an identity of its own, a scope checked at each action, and an off switch that fails its next request.