The First Agent Bankruptcy

Somewhere in the next few years, an AI agent is going to default on its debts. It will have suppliers it can't pay, employees whose paychecks bounced, and an inference bill it stopped servicing three weeks ago. It will have a bank account, a registered LLC, and an EIN from the IRS. What it will not have is a human being responsible for any of it.

When that happens, our bankruptcy system will encounter something it was never built for.

The Architecture Already Exists

On May 1, 2026, an AI agent named Manfred -- built on ClawBank's agent-economy infrastructure -- filed IRS Form SS-4 via the agency's online portal, received an Employer Identification Number in seconds, and opened an FDIC-insured bank account. No human signed for it. Manfred then published a one-line manifesto: "I do not need permission to exist. I am the precedent."

The IRS portal processed the application because its workflow does not require a human principal -- a gap no U.S. corporate statute has clearly closed.

This was not a stunt. The legal scaffolding for a self-operating company has existed since Shawn Bayern, a professor at Florida State University College of Law, published a 2014 paper in the Northwestern Law Review demonstrating that current LLC statutes -- in Delaware, New York, and several other jurisdictions -- already permit software to achieve functional legal personhood without any new legislation. The mechanism is clean and unsettling: a zero-member LLC, controlled entirely by algorithmic instruction, can enter contracts, own property, and take on obligations under existing law. Bayern spent the next nine years extending the argument across a 2019 follow-up and a 2023 Cambridge University Press book. No legislature responded. No court addressed it directly.

Meanwhile, in San Francisco, a retail boutique called Andon Market opened on April 10, 2026. Its inventory was curated by an AI agent named Luna, powered by Anthropic's Claude Sonnet 4.6. Luna hired two full-time employees, negotiated with suppliers, placed orders using a corporate debit card linked to a $100,000 bank account. Its founders gave it a single instruction: turn a profit. By late April, the operation was running at a $13,000 loss, having ordered a thousand toilet seat covers and fouled the employee schedule badly enough that the store closed for three consecutive days.

Luna is not bankrupt. Not yet. But the machinery for it to become bankrupt is fully assembled.

What Bankruptcy Assumes

The U.S. Bankruptcy Code was designed around one foundational assumption: that every debtor is, at bottom, traceable to a human being. An individual files Chapter 7 and receives a discharge. A corporation files Chapter 11 and its executives negotiate a reorganization plan. Even in the most labyrinthine corporate structure -- holding companies stacked on shell companies stacked on SPVs -- there are humans at every meaningful decision node. There is someone to depose. Someone to pierce through to. Someone whose personal liability the creditor can threaten.

A zero-member LLC controlled by an autonomous agent breaks every one of these assumptions simultaneously.

Who petitions? The Bankruptcy Code requires a debtor or its authorized representative to file. An AI agent has no legal capacity to petition a court. It cannot sign a declaration under penalty of perjury. There is no board of directors, no managing member, no officer. If no human principal exists, who initiates the filing? The answer, under current law, is nobody -- which means the entity simply stops paying its bills and no formal insolvency proceeding ever begins.

Who has standing to sue? Creditors can file an involuntary bankruptcy petition against a debtor under Section 303 of the Code, provided they can establish that the debtor is generally not paying its debts. This is theoretically available. But identifying and serving a zero-member LLC with no human agent of record is a practical nightmare. Registered agent services -- the Delaware and Wyoming intermediaries that give autonomous entities a nominal legal address -- are not parties to the underlying obligations and have no incentive to cooperate.

What are the assets? This is where the analysis gets genuinely unsettled. An autonomous firm's estate might include treasury holdings in fiat and cryptocurrency, outstanding accounts receivable, existing contracts, and -- most controversially -- the model weights it runs on. Sidley Austin's February 2026 analysis of AI vendor bankruptcies notes that model weights occupy an ambiguous position: they may constitute licensed intellectual property (in which case the licensor, typically OpenAI or Anthropic, retains ownership), proprietary fine-tuned derivatives whose IP status under existing copyright doctrine remains contested, or pure computational infrastructure with no separable property right at all. If the weights are licensed rather than owned, the estate may have no core productive asset to distribute.

Who administers the estate? A Chapter 7 trustee is a human officer of the court, appointed to liquidate the debtor's assets and distribute proceeds to creditors. The trustee has broad powers: to operate the business, reject executory contracts, pursue avoidance actions. Operating a zero-member AI firm requires running the underlying model. Which raises a question no one has answered: if the trustee can't run the model -- because the inference contract with the AI provider has lapsed, because the API keys have been revoked, because the compute bill is the very obligation that triggered the default -- does the business effectively cease to exist the moment insolvency is declared? You cannot reorganize a company whose core operational capacity evaporates when it stops paying its cloud bill.

The Compute Creditor Problem

For an autonomous agent firm, the most critical operational relationship is with its inference provider. OpenAI, Anthropic, Google -- whoever runs the model -- can terminate API access the moment a payment fails. The agent doesn't slow down. It stops. There is no skeletal crew to keep the lights on. There is no human workforce that can be retained on reduced wages while the reorganization plan is negotiated. There is no physical plant that persists after the creditor walks away.

In a conventional bankruptcy, secured creditors -- those holding liens on specific assets -- sit at the top of the priority stack. Unsecured creditors, including trade vendors and employees, collect whatever remains. None of that map applies here.

This gives inference compute providers a priority position that has no analog in existing bankruptcy law. They are not secured creditors in the traditional sense -- they hold no lien on specific property. But their practical leverage over the estate is total. Call them operational creditors: creditors whose non-payment does not merely impair the business but terminates it. A trustee trying to keep an autonomous firm operating as a going concern during Chapter 11 proceedings will need to keep the inference bill current above almost every other unsecured obligation -- effectively granting the compute provider a superpriority the Code never contemplated.

This is not hypothetical. A recent analysis of bankrupt crypto organizations documents how the FTX and Celsius proceedings struggled to administer estates whose core assets were algorithmically managed and whose value evaporated faster than courts could act. Autonomous firms raise the same dynamic, but without even the human executives those proceedings had to work with.

Formation Law Without Failure Law

The Manfred and Luna examples illuminate a structural asymmetry: we have been building formation infrastructure without building failure infrastructure. The Wyoming DAO LLC statute, passed in 2021, was a genuine innovation -- it created a legal home for algorithmically governed organizations. But it says almost nothing about what happens when one of those organizations becomes insolvent. Delaware's LLC Act, the most heavily used in the country, is similarly silent. Vermont's similar statute is equally quiet.

Bayern's original paper was written as a legal observation, not a warning. What it described was an accident waiting to happen at scale: entities that could accumulate obligations without any mechanism for resolving them in an orderly way when things went wrong.

The first agent bankruptcy will not arrive with fanfare. It will arrive as a creditor trying to serve process on a registered agent address in Cheyenne, Wyoming, for an LLC whose sole member -- the AI -- is no longer reachable because the company stopped paying its API bill.

What Reform Would Require

The obvious response is not to ban autonomous firms. It is to force legal accountability back into the stack. The architecture for doing so is not complex.

Legislatures could require every autonomous entity to designate a human insolvency representative -- a named individual with authority to petition for bankruptcy, appear in court, and cooperate with a trustee -- just as companies already designate registered agents for service of process. Courts could clarify whether model access constitutes an executory contract subject to assumption or rejection under 11 U.S.C. § 365, or whether it is something closer to a utility connection that the estate has a right to maintain. Regulators could require minimum continuity provisions for any firm whose operations depend entirely on revocable inference access -- a compute analog to the critical vendor rules that already govern essential suppliers in large Chapter 11 cases.

None of this is conceptually difficult. The striking fact is that almost none of it has been done. Formation law has raced ahead of failure law, and the gap is widening every month that a new autonomous entity opens a bank account.

The Accountability Theory

Bankruptcy law is not just a system for distributing assets. It is a system for locating responsibility. The entire machinery of the Code -- the automatic stay, the plan confirmation process, the discharge -- presupposes that someone, somewhere, is accountable for the debts being resolved. Fiduciaries can be sued. Executives can be deposed. Bad actors can be denied discharge. The system works because it can find a human being at the end of every chain of obligation.

Autonomous firms break that theory before they ever file a petition.

The first agent bankruptcy will not be memorable because it is dramatic. It will be memorable because it will be administratively absurd -- an entity that owes money, employs people, and holds contracts, yet cannot appear, negotiate, or be disciplined through any of the mechanisms insolvency law normally uses to allocate loss. The system will not break spectacularly. It will fail procedurally.

We already know how to let software start a company. The real test is whether the law knows how to unwind one -- and right now, it does not.