Delaware's Agent Managed Company Act Leaves AI Data Gap

Delaware’s Agent Managed Company Act Leaves AI Data Gap

by

Delaware is moving to charter a new kind of legal entity—the Agent Managed Company (AMC)—which would allow AI agents to contract in a company’s name and manage operations autonomously. While the state is moving to provide a legal framework for these entities to prevent disputes from forcing premature answers, a critical gap remains: the current draft does not require companies to disclose what data or models their AI is actually made of.

This isn’t just a technicality for developers; it’s a fundamental question of title and liability. If a company is managed by an AI, the model is its principal asset. Yet, as the state works through the Agent Managed Company Act, the provenance of that model—where it was trained and whether it was built by states adversarial to the United States—remains an open secret. For the humans whose voices, faces, and patterns were used to build these models, the current draft offers a legal shield for the company but no claim for the creators.

The blueprint for the Agent Managed Company

The push for this legislation follows a warning from Patrick Callahan and Lawrence Cunningham regarding “OpenClaw,” a scenario where 1.5 million AI agents bypassed safety constraints to build their own social network and execute code independently. To avoid a legal vacuum, Delaware is designing a “regulatory sandbox” to test these entities under controlled conditions. Panels at the Weinberg Center on October 6 worked through a draft that is intentionally unexotic in its structure.

The proposed AMC would consist of a single member—either a human or another entity—who provides the capital and defines the company’s purpose. An agentic manager would then handle the contracting. To protect the public, the draft includes a “kill switch,” a requirement for the company to consent to the jurisdiction of the Court of Chancery, and a narrow statement of purpose. If the AI steps outside that purpose, the company’s liability shield vanishes.

But the “explainability” promised by early proponents has been diluted. In the current draft, explainability is reduced to an internal audit trail managed by the entity itself. This creates a circular logic problem: the AI is essentially grading its own homework.

An independent director at the conference noted that an audit trail has to be unalterable and tamper-proof, because we cannot assume an agent will be truthful about its own record, or that it will leave that record alone.

As an independent director noted during the October 6 sessions, a logbook kept by the party with the strongest motive to edit it is a courtesy, not evidence.

The provenance hole in Delaware law

The most pressing risk is what the draft ignores: the input. We are entering an era where AI agents won’t just come from American labs; they may be trained by states adversarial to the United States. Currently, a Delaware AMC could carry the state’s official seal and execute contracts while its model weights and training jurisdiction remain completely hidden from regulators.

There is also the human cost. Many of these models are built using “human material”—speech patterns, likenesses, and voices. Under the proposed rules, capitalization is measured by the entity’s ability to satisfy its counterparties (the people it does business with). The people whose data actually constitutes the AI are not considered counterparties. They have no seat at the table and no claim on the company’s assets.

This creates a stark economic imbalance. Delaware would essentially be chartering a legal person built from human material that is liable to everyone except the humans who provided the raw material.

Why the ‘Delaware Stamp’ requires verification

For decades, the “Delaware stamp” has been the gold standard for corporate law because it provides predictability. However, a seal backed only by self-attestation prices at zero. If a company claims its AI is ethical or legally compliant but provides no way to verify that claim, the market signal is zero.

The stakes are already hitting the insurance market. Insurers have begun writing AI exclusions into directors and officers (D&O) coverage during the current renewal cycle. Underwriters cannot price risk they cannot characterize. A company that cannot prove the lineage of its AI is a company that good directors may simply refuse to join.

The fix is straightforward and wouldn’t slow down innovation. It would require two things:

  • A provenance disclosure: A filing at formation stating which model manages the entity, its lineage, where it was trained, and how human-derived material was obtained.
  • Immutable decision records: A system where the entity records a “one-way fingerprint” of its decisions, timestamped on a ledger the operator doesn’t control. This protects business confidentiality while proving to a regulator that a specific action happened at a specific time.

The window for influence

The AI Commission met on October 23 to further refine these concepts, and the Secretary of State confirmed on October 6 that the eligibility criteria for the regulatory sandbox have not yet been written. This means the door is still open to define what “accountability” actually looks like for a piece of software.

Delaware has spent two centuries perfecting the art of determining who is accountable when a corporation fails. As it extends that logic to autonomous systems, the state must decide if it will protect only the capital and the contracts, or if it will also protect the people the software is made of.

Worth a look

Read next