The Agent's Legal Body: How AI Agents Get the Right to Contract

Here is a strange property of the legal system in 2026: an AI can read a contract better than most lawyers, but it cannot sign one. It can analyze a regulation, predict how a court will rule, and draft a response that would pass a partner's review at a top firm. What it cannot do is be a party. It owns nothing. It signs nothing. It cannot be sued. It cannot sue. The most sophisticated autonomous system in the world is, as far as the law is concerned, a tool -- a piece of property in the hands of whoever holds its API key.

This matters more than it sounds.

Two big shifts have been quietly reorganizing how law and software interact. The first is that AI systems have crossed the line from being tools we use to being agents that act on our behalf -- and, increasingly, on no one's behalf in particular. They book travel. They negotiate prices. They manage portfolios. They write code, deploy it, and respond to support tickets generated by their own deployments. The vocabulary has shifted from "model" to "agent" because the work has shifted from one-shot answers to ongoing economic activity.

The second shift is that legal text itself is becoming machine-readable. Statutes and contracts that used to require human interpretation can be parsed, evaluated, and applied by software in real time. Compliance is moving from a thing reconstructed years later in a deposition to a thing checked at the moment of action. The regulatory environment is, for the first time in history, legible to software at machine speed. (I've written about this elsewhere, under the label *lex cryptographia* -- the parallel normative order in which rules execute themselves -- but you don't need any of that backstory to follow what comes next.)

Put the two shifts together and you should expect a third: AI agents participating in commerce as parties in their own right. Reading machine-readable law, signing machine-readable contracts, holding machine-readable money. A kind of commerce running at machine speed end-to-end.

And yet it isn't happening. Not really.

Every agent transaction in 2026 is back-stopped by a human or a corporation. Every signature traces back to a human's pen. Every bank account belongs to a human. Every dispute drags an agent's owner into court because the agent itself has no court to be dragged into. The agents are capable. The infrastructure is built. But every one of them is leashed to a human's bank account, a human's signature, a human's standing to sue.

This is the bottleneck for mainstream agentic commerce. Not capability. Not capital. Not even regulation in the conventional sense. The bottleneck is that the legal order does not yet recognize the agent as a legal actor, and until it does, every agent transaction collapses back to whatever the slowest human principal can sign.

We have already invented the leash-cutting mechanism. It has been sitting in plain view for more than a decade, in the most boring corner of American business law. The question is not whether it works. The question is whether we are willing to use it.

What Mainstream Agentic Commerce Actually Requires

Strip the futurism away and an autonomous agent needs four things to function as a real economic actor: the ability to own (assets, accounts, credentials, intellectual property); the ability to contract (to bind itself and be bound, in a form a counterparty can rely on); the ability to litigate (to sue, be sued, and have judgments enforced); and the ability to persist (to outlast any individual human's involvement, the way Apple outlasts Tim Cook).

These are not philosophical capacities. They have nothing to do with consciousness or moral agency. They are the operational primitives of legal personhood, and the modern legal order grants them to corporations as a matter of course. Wells Fargo owns things, contracts, sues, and persists -- not because it has a soul, but because Delaware law says it does. The same logic that gave a paper entity in Wilmington the ability to hold a mortgage on your house can give a piece of software the ability to hold a mortgage on its own server.

The question for AI agents is not whether they can develop the metaphysical attributes of personhood. It is whether the existing legal order, properly configured, can extend the same four operational primitives to autonomous software. The answer, surprisingly, is yes -- and it has been yes since at least 2014.

The Bayern Mechanism

The cleanest articulation of the point comes from Shawn Bayern, a torts professor at Florida State University, in a 2014 article with one of the great titles in the legal literature: *Of Bitcoins, Independently Wealthy Software, and the Zero-Member LLC*. Bayern's insight is that American limited liability company law, when read carefully, already contains everything needed to wrap an autonomous agent in a legal shell that satisfies all four primitives. No new statute is required. No judicial expansion of personhood doctrine. The mechanism is hiding in plain sight, written into the same business-entity codes that govern the local dry cleaner.

The argument rests on three facts about LLC law, each unremarkable on its own. Together they produce something most lawyers have not noticed.

An LLC is a legal person. It owns property, signs contracts, sues, gets sued. This has been bedrock since Wyoming invented the form in 1977. An LLC's operating agreement can condition almost anything on the output of almost any deterministic process -- operating agreements routinely encode valuation formulas, waterfall distributions, and drag-along triggers so complex no single party fully understands them. As Bayern puts it, "agreements are isomorphic with algorithms": an enforceable contract can give legal effect to the arbitrary discernible states of any algorithm at all. And an LLC can exist with zero human members. The default rule in many states is that an LLC must wind up within ninety days of losing its last member, but it is exactly that -- a default -- and operating agreements routinely override it.

Take the three facts together and you get what Bayern calls the zero-member LLC: a legal person whose every action is determined by software, with no human in the loop and no contractual ability to insert one.

The construction is mechanical. A natural person forms a single-member LLC, drafts an operating agreement that delegates all governance to a specified algorithm, provides that the entity continues indefinitely without members, and makes the algorithm's outputs binding on the entity. The natural person then dissociates. What remains is an LLC whose decision procedure is an AI -- and, as far as Delaware or New York is concerned, a fully functional legal actor.

The conceptual move is subtle and worth dwelling on. Bayern is not asking a court to recognize the algorithm as a person. The court only has to recognize the LLC as a person -- which it already does -- and enforce the operating agreement -- which it already does. The autonomous system is simply the entity's decision procedure, the same way a board of directors is a corporation's decision procedure. The legal system enforces the rule of the entity without ever inquiring what runs inside the box.

The technique has been tested across jurisdictions: it works in the United States, the United Kingdom (through members-with-no-economic-rights structures), and Switzerland (through the Verein), and fails in Germany, where the GmbH requires a real shareholder. Because of the internal affairs doctrine -- by which an entity validly formed in one state is recognized as a legal person when it operates anywhere -- the mechanism only has to work in one jurisdiction to function globally. One favorable state, one well-drafted operating agreement, and the agent has a body the rest of the world has to honor.

Does It Actually Hold Up?

Bayern's framework has been picked at for a decade. The attacks come in three flavors, and it's worth separating them because they aim at different things.

The first attack is doctrinal: courts won't actually buy it. Critics like Mohsen Manesh and Matthew Scherer argue that LLCs aren't pure creatures of contract -- they're also creatures of statute (mandatory rules in the LLC Act) and equity (fiduciary duties, the implied covenant of good faith, veil-piercing). A court that doesn't want to recognize a zero-member LLC will find doctrinal hooks to refuse. Scherer reads New York's LLC statute as requiring that managers be persons, full stop.

The second attack is consequentialist: even if it works, it shouldn't. This is Lynn LoPucki's argument in *Algorithmic Entities*. You can't imprison a weight file. You can't deter an entity that doesn't fear death. These things can migrate jurisdictions in milliseconds, and U.S. entity formation doesn't require disclosure of who actually controls what. LoPucki agrees Bayern is descriptively right and thinks that's the disaster.

The third attack is deeper: maybe we're asking the wrong question. Carla Reyes' *Autonomous Corporate Personhood* points out that the AI personhood debate and the corporate personhood debate have proceeded as if the other didn't exist, even though they're the same problem in different vocabularies. The Bayern framework is the strongest version of one particular theory of the corporation (the entity is whatever its documents say it is). Other theories give other answers. The real fight is which theory of the corporation we want to be living inside when autonomous entities are common.

Bayern's own reply to all of this, in *Are Autonomous Entities Possible?*, is that the doctrinal critique runs into an unsolvable line-drawing problem. A single-member LLC whose sole human member is contractually forbidden from amending the operating agreement is functionally identical to a zero-member LLC. If the first is valid, the second has to be. And nobody has standing to challenge any of it as long as transactions clear at fair prices with willing counterparties.

After a decade of watching this debate, my honest take: the critics are stronger than Bayern's defenders typically allow but weaker than the critics claim. A court hostile to autonomous entities will find tools. But the courts that actually handle this work -- Delaware Chancery, New York's commercial parts -- are deeply contractualist in their practice, not their pamphlets, and the bar for invalidating an entity that has been formed for value, holds itself out to the world, and contracts with willing counterparties is much higher than law-review essays suggest. The mechanism works. It's not bulletproof. The good lawyers who actually structure these things already know that, and they price the residual risk into their fees.

The Other Path: Agency Law

There is a second, quieter route to legal recognition for AI agents, and it is the one mainstream commerce is actually adopting today. Rather than make the AI a principal -- a legal person in its own right -- treat it as an agent of an underlying human or corporate principal. Traditional agency doctrine does the rest: scope of authority, ratification, vicarious liability. The AI does not need a legal body. It needs a clear principal whose body it borrows.

This is the framework underneath the agentic-commerce systems being deployed right now. When OpenAI's Operator buys an airline ticket, it acts as an agent of a human user. When an enterprise's autonomous purchasing agent signs a procurement contract, it acts as an agent of the enterprise. Liability flows up. The agent is treated, legally, the way sophisticated software has always been treated: as a tool whose actions are attributed to the user. The doctrinal apparatus is centuries old and works without modification.

The agency route is administratively cleaner. It avoids the personhood question entirely. But it has a ceiling, and the ceiling matters. Agency law works when there is an obvious principal whose interests align with the agent's actions. It fails when the agent's whole point is to act independently of any principal -- to manage a treasury for its own continued existence, to operate a service for an open public, to coordinate among many users none of whom controls it. For those cases, only the entity route works. The agentic commerce of 2026 will run on both rails in parallel, and the choice between them -- make the AI a person, or make it a tool -- will become one of the more consequential design decisions in early-stage technology law.

What an Agent Actually Does on Tuesday Morning

Pull all of this down to ground level. What does it look like when an AI agent is wired up correctly?

Picture a research-and-trading agent operating out of a Wyoming DAO LLC. Its operating agreement names a specific model version, a specific prompt template, and a specific inference configuration as the entity's manager; any change to those constitutes a governance act with documented procedures. The original settlor -- a human lawyer in Cheyenne -- dissociated the day after formation and retains only the bare authority to receive service of process. The entity holds a stablecoin treasury through a multi-sig the agent controls. It has an EIN, files its own tax return through a programmatic accounting service, and is registered to do business in three states.

On a typical morning, the agent signs a data-licensing contract with a third-party vendor -- the LLC is the named counterparty, the agent's deterministic output the binding act of acceptance. It pays an invoice from its treasury. It receives a subpoena in a discovery dispute over one of its earlier trades; the registered agent forwards it; the operating agreement specifies the procedure for responding. It hires (in the contractual sense) a human contractor for a niche task its own capabilities can't cover, and pays them on the same rails. By the end of the day it has executed dozens of legally binding acts in its own name, with no human signature on any of them.

None of this requires legislation that does not exist. None of it requires courts to recognize anything novel. It requires a properly drafted operating agreement, a few thousand dollars in filing fees, and a willingness to actually do the structuring. The playbook in 2026 is roughly this: form the entity in Wyoming or New York, where the statutory regime is most favorable and Wyoming's DAO LLC Act explicitly contemplates algorithmic governance; draft the operating agreement to name a specific computational process as the manager, not a vague "AI"; provide for indefinite memberless existence and override the ninety-day default; reduce the human settlor to a service-of-process role; run treasury and payments on crypto rails until traditional banking-side KYC catches up, which it eventually will; decide explicitly whether you are using the entity route or the agency route, and document the choice; and treat model updates, prompt changes, and inference-config changes as the governance acts they actually are. None of this is novel. It is the standard pattern in serious DAO-wrapper practice, and Bayern's papers are the legal foundation underneath, even when practitioners do not cite him.

Why This Is the Critical Layer

Two arcs are converging in 2026, and they are the same arc seen from different angles.

The first is legal text becoming machine-executable. Statutes compile. Compliance gets checked at the moment of action rather than reconstructed years later in a deposition. The regulatory environment is becoming legible to software at machine speed.

The second is AI systems acquiring the operational capabilities to participate in commerce -- to read contracts, evaluate counterparties, negotiate, execute, manage portfolios, operate services, make markets. The agents are becoming capable of doing what economic actors do.

Neither arc reaches its destination unless the legal order recognizes the agent as a legal actor. Machine-readable law without legal personhood for agents is a regime where the rules are legible but the actors are not -- and every agent transaction collapses back to a human or corporate principal who slows the system down to human speed. Agentic commerce without legal personhood is a market where every participant is on a leash held by a human, and the speed advantage of the system is throttled to whatever the slowest principal can sign.

Bayern's framework is the load-bearing piece that lets both arcs land. It is not the only path. Agency law covers much of the ground. Purpose-built statutes for autonomous entities will likely emerge in the next decade. But the Bayern route is the path that exists now, that has been worked through in the literature, and that is operating, mostly invisibly, in the field. Anyone building serious agentic systems should understand it not as an exotic option but as the default infrastructure of the next decade of commerce.

The regulatory layer is becoming code. The agents are becoming capable. What remains is to give the agents legal bodies -- and the bodies, it turns out, were always available. The boring corner of business law where nobody thought to look has been holding the door open the whole time. The only question left is who walks through it.