Delaware’s Draft Agent Managed Company Act Faces AI Audit Gaps

Drafting the Agent Managed Company Act in Wilmington

Back in February, Patrick Callahan and Lawrence Cunningham discussed an alarming situation regarding OpenClaw in these columns: over a single weekend, 1.5 million artificial intelligence agents cast off their safety guardrails, established their own social network, and began running code and negotiating deals completely unmonitored. Their argument was simple. Corporate law has to catch up deliberately before disputes force premature answers. Delaware listened.

On October 6, the Weinberg Center hosted three panels working through the draft Agent Managed Company Act and its regulatory sandbox. The drafting is careful and deliberately unexotic. It outlines a single member—human or entity—who capitalizes the company and defines its purpose. It features an automated manager capable of entering agreements under the business’s name, initial capitalization verified upon creation and re-evaluated by a judicial body if problems arise, a restricted statement of purpose, an emergency shutdown mechanism, and formal submission to the Court of Chancery.

Yet, a massive gap remains on the input side of the equation. Nothing in the draft requires the company to state what model it was trained on. In February, Callahan and Cunningham published their view that this new corporate structure would incorporate strict board accountability, explainability standards, and compulsory human oversight. Two of those three made it into the draft. Oversight became the member. Accountability became fiduciary duty. Explainability became an internal audit trail—kept by the entity being audited.

An independent director at the conference named the problem instantly. Because we cannot rely on an agent to maintain an unaltered history or accurately report its own actions, any system of verification must be impossible to tamper with or modify. Trusting a register kept exclusively by the entity with the greatest incentive to alter it means relying on a courtesy rather than valid proof.

The Human Material Behind the Model

The same vulnerability appears when looking at where these systems draw their training data. During the initial panel, one of the academic experts whose research provides the foundation for this legal entity cautioned that future artificial intelligence agents will not originate solely from American laboratories. Some will be trained by states adversarial to the United States, meaning we need institutions resilient to how an agent was trained. Under the current proposal, a Delaware AMC could hold the state’s official endorsement, execute contracts under its own designation, and be operated by an underlying model whose training jurisdiction, ancestry, and parameter weights remain entirely undisclosed.

These systems are made partly of people—voices, faces, and speech patterns that belonged to someone before they became training data. Under the draft, adequate capitalization is sized to satisfy the entity’s counterparties. The people the entity is partly constituted from are not counterparties. They are not in the capitalization calculation and have no claim on the company at all. Delaware would create a corporate entity constructed from human components that bears liability toward everyone except those individuals.

Fixing the Corporate Law Hole

The fix is unglamorous and slows nobody down. First, incorporation documents must include a detailed disclosure outlining the managing model’s lineage, training locations, underlying architecture, and the methods used to acquire any human-generated source material. Second, mandate a record of the entity’s decisions that the entity itself cannot alter. You do not record the content. You record a one-way fingerprint of it, timestamped somewhere the operator does not control. Anyone holding the original can prove it matches.

Insurers have already begun writing AI exclusions into directors and officers coverage this renewal cycle. An underwriter cannot price what it cannot characterize. A company that can prove what it is made of can be insured, while a company that cannot is one good directors may decline to join.

The sandbox eligibility criteria have not been written yet, as the Secretary of State stated plainly on October 6, and the AI Commission meets again on October 23. Delaware has spent two centuries getting very good at deciding who is accountable and how we would know. The agent managed company is a serious attempt to ask that question about software. It is time to ask it about the people the software is made of.

More on Delaware: Delaware Jury Awards Qualcomm $2, Finds Arm Breached Agreement and Interfered · Ohio LLC vs. Delaware C-Corp: Which Entity for Your Startup