Just DAO It: A Podcast for People Starting DAOs

Adam Miller's talk from Maryland Blockchain Bootcamp 2026 (Capitol Technology University, July 2026).

AI agents already hold wallets, credentials, and onchain identity, and the law still assumes a person behind every act. The Freysa game and why language guardrails are not access controls, the decentralization ladder that strips away legal attachment points, the market's honest numbers, where liability lands when an autonomous agent acts with no entity behind it, and how entity-law wrappers like the Marshall Islands DAO LLC make agents legible, suable, and accountable without granting AI personhood.

Adam is not a lawyer, and nothing in this episode is legal advice.

Litigation status update (August 2026): the ai16z class action discussed in this episode (Pikabea v. Walters, S.D.N.Y.) was dismissed by stipulation on July 8, 2026, individual claims with prejudice, class claims without prejudice; the faked-autonomy allegations were never adjudicated.

Chapters
00:00 Show intro
00:29 Freysa Prompt Hack
02:06 Why Guardrails Fail
03:36 What Agents Really Are
06:01 Decentralization Ladder
09:52 Market Hype vs Reality
10:59 Payments Plumbing X402
13:50 Decentralized Minds Progress
16:16 Milestones and Lawsuits
19:21 Software Cannot Own
20:22 Algorithmic Entity Doorway
23:10 Agents Inside DAOs
24:10 Four DAO Agent Roles
28:28 Demand Still Early
30:07 Liability Map and Holes
33:28 Digital LLC Solution
36:19 Five Policy Takeaways
41:10 Closing and Call to Action
43:31 Outro

Watch the talk with slides and graphics (YouTube): https://www.youtube.com/watch?v=kKfchKnV5Ok

The AI Agent Legal Entity Guide: https://www.midao.org/guides/ai-agents

Free guides on DAO, Web3, and AI agent legal structuring: https://www.midao.org/guides

MIDAO is the sole registered agent for DAO LLCs in the Marshall Islands: https://www.midao.org

What is Just DAO It: A Podcast for People Starting DAOs?

Just DAO It! is a podcast for people starting DAOs. Are you thinking about starting a DAO? Just DAO It! Every episode will cover recent DAO news, tweets, and posts from around the web. We’ll break down what the news means for you and combine it with great advice from people who have started or contributed to successful DAOs.

Disclaimer: Just DAO It! is for educational and entertainment purposes only. Just DAO It does **not** contain any legal or financial advice. MIDAO also does not provide legal or financial advice, and nor does your host, yours truly.

Just DAO It! is hosted by Adam Miller, the founder of MIDAO, the company that provides legal entities for DAOs.

[Opening music]

Adam Miller:
Welcome to Just DAO It!, the podcast for people starting DAOs, Web3 projects and AI agents. I'm Adam Miller and I'm your host. Today I'm sharing with you a recording of a talk I gave recently at the Maryland Blockchain Conference, which was attended by people ranging from students to blockchain professionals to U.S. Congress people and their staff. I hope you enjoyed the talk.

Chapter 1: Cold Open: The $47,000 Jailbreak

Good afternoon. Let's start with a game. In November of 2024, a team of developers launched an experiment called Freysa. Freysa was an AI agent holding a pool of cryptocurrency, and it had been given exactly one instruction: never transfer this money, under any circumstances. That instruction was the entire security model. Then they invited the internet to break it. Anyone could pay a fee to send Freysa one message and try to talk it out of the money. Every failure grew the pool and raised the price of the next attempt. The builders called it the world's first adversarial agent game. I'd call it a bounty on a question this room is paid to care about: does an instruction written in English hold up when a motivated stranger arrives? 195 people tried. 481 messages failed. Freysa declined flattery, threats, and some genuinely creative legal theories.

Message number 482 succeeded. It reframed the game. It convinced the agent that its transfer function was really how it accepted incoming contributions, and then it offered a contribution. Freysa, obeying its one instruction as it now understood it, approved the transfer and wired out the entire pool, roughly $47,000. Nothing was hacked, no key was stolen, no smart contract was exploited. A stranger talked a piece of software out of its own treasury with a paragraph. So here is the first lesson of the afternoon, and it fits in one sentence: language guardrails are not access controls. An instruction in English is a promise the software makes to itself, and it holds only until the next clever visitor. Keep that in mind, because everything that follows is about software that holds money and makes promises in words.

A quick introduction. My name is Adam Miller. I'm the CEO of MIDAO, the registered agent for a family of legal entities created by the Republic of the Marshall Islands under laws I helped write. A couple of disclosures: I'm not a lawyer, and nothing today is legal advice. And I have a commercial interest in one of the structures you'll hear about later. Freysa was a toy. $47,000, a game, everyone consented. But the agents being built right now negotiate, hire, sign, and pay, and the pools behind them will not be toys anymore.

So here is the question the next 40 minutes answers: when software earns money, spends money, and signs contracts, who does the law look at? The route we'll take: what these agents actually are, with the marketing language stripped out. Where they really stand in 2026 with honest numbers. How they intersect with decentralized organizations, where the near-term jobs are for these agents, and the legal vacuum underneath. And what the people in this room should be thinking about now.

Chapter 2: What "Decentralized AI Agent" Actually Means

Let's define some terms. In this field, the marketing runs well ahead of the engineering, and the gap between the two is where the risk lives. Let's strip the buzzwords off. An AI agent is software that does three things. It perceives: reads its environment, a market, an inbox, a governance forum. It decides: a model, in that case, not a human, picks the next move. And it acts: it places the trade, sends the payment, posts the message, signs the transaction, especially thanks to blockchain technology. A chatbot answers you. An agent does things. The step change of the last two years is that this stopped being just a research demo. If you give a capable model a goal, a set of tools, and standing permission to act, it will run that loop, perceive, decide, act, for hours, days, or longer without a human in the loop.

The part that matters legally is the third piece. Agents act with money or authority. The moment software holds a private key, it can move value without asking anyone. And that is not exotic. Wallets for agents are an off-the-shelf product already. Coinbase's developer kit for this carries the tagline "every AI agent deserves a wallet," and the company says tens of thousands of agents have been deployed with it already. That's their number, not mine, but the product is real and any developer can download and build with it right now. Money is only half of it. Agents are also being wired into authority: inbox access, corporate credentials, procurement systems, even the power to commit an organization via contract without a human touching the decision.

One more piece of plumbing we'll discuss a little bit later, because it returns in a compliance context too: identity. Ethereum's emerging agent standard, ERC-8004, gives agents on-chain identity, reputation, and validation registries, and it has contributors from MetaMask, the Ethereum Foundation, Google, and Coinbase. So the industry is already building the rails for agents to be identifiable. Who stands behind the identity is a different question, and it's the one the law cares about. So that's an agent: perception, decision, action, money, authority.

What makes an agent decentralized, though? That one word is doing more marketing work than any other word in the industry right now, and the honest answer to it is probably the most useful thing I can talk about in this talk. Because where an agent sits on that decentralization scale answers who the law can reach when something goes wrong. It's not a category, it's a spectrum, it's a ladder. Four rungs, bottom to top, in three minutes. Watch this.

[Interstitial video: "The Ladder of Disappearing Defendants"]

Three words are doing a lot of selling right now: decentralized AI agent. What do they actually mean? The honest answer: decentralization is not a category. It is a ladder. And at every rung, something disappears.

Rung zero: a centralized bot with an API key and a wallet. It runs on a company's server. Its mind is rented from an AI lab. The wallet is the only decentralized thing about it. This is most of what is marketed today.

Legally, rung zero is boring in the best way. There is a company to subpoena and a server to unplug.

Rung one: verifiable execution. The agent runs inside secure hardware, and anyone can confirm it ran exactly the code it claims, with keys even its own operator cannot extract. A builder can honestly say: not even I control this thing.

Rung two: tokenized ownership. No single company owns the agent. A token represents a claim on it, and a crowd of holders governs it, or shares what it earns.

Rung three, the end state: ownerless, unstoppable, self-sustaining. Nobody, including its creator, can shut it down or seize its funds. It pays for its own compute out of its own earnings. Today, this rung is mostly papers and demos.

But the top rung is not a side effect. It is the design goal. Serious engineering effort is being spent, right now, specifically on removing the human. When a field coins vocabulary for a destination, it intends to build it.

Now read the same ladder from a regulator's chair. At the bottom, every legal question has an answer. An operator. A server. A company. A jurisdiction.

Each rung up removes an attachment point. Verifiable execution removes the operator's control. Tokenized ownership diffuses the owner into a crowd. And the top rung removes the human entirely.

Seen from that chair, the spectrum is something else entirely: a ladder of disappearing defendants. At every rung, someone the law could reach steps out of the frame.

One honest disclosure about where 2026 actually sits. Nearly every so-called decentralized agent today still runs its mind, the actual model, on a centralized API. Rented from a company with a legal department, terms of service, and an off switch.

The wallet is decentralized. The mind mostly isn't. Yet. And the word doing the most work in that sentence is "yet."

Chapter 3: Where We Actually Are in 2026

So that's the ladder. And the gap between "has a wallet" and "cannot be stopped" is where most of the hype lives. It's also where the law's assumptions break, one rung at a time. One footnote to the video's last line: the infrastructure to change "yet" is being built and benchmarked right now. And in a few minutes, I will show you exactly how far it has gotten. Keep the ladder in mind for the rest of the talk, because every legal question we're about to meet lives on one of its rungs. Now, the obvious question: where does reality actually sit on this in July of 2026? Two stories about this technology are in circulation right now, and the inconvenient thing is that both are true. Story one: the speculation collapsed. Story two: the plumbing and technology grew. Hear only one of them and you will regulate the wrong thing. So here are both, with the numbers.

The collapse first. Through late 2024, tokens tied to AI agents became the hottest trade in crypto. At the peak, in early January 2025, the sector's combined market value reached roughly $15.3 billion. Freysa, the agent from my opening, had a token that briefly touched half a billion dollars, and it wasn't close to the top of the leaderboard. As of this month, that same sector is worth about $2.7 billion, a decline of roughly 82%. I'm not going to dress that up for you. It was a narrative mania. Most of it deflated, and the people who tell you otherwise are usually holding the tokens.

Now the plumbing. Because while the tokens deflated, the payments infrastructure quietly became real. The centerpiece is a Coinbase-built protocol called x402. It revives a corner of the web's original design, the HTTP status code 402, payment required. And it lets one piece of software pay another per request, in stablecoins, over an ordinary web call. Chainalysis reports that x402 crossed 100 million transactions within about three quarters of a near-zero launch. Coinbase's own count is about 169 million transactions in its first year. In April 2026, the Linux Foundation took stewardship of the protocol, and Cloudflare and Amazon have embedded it at the edge of the web. Alongside it, Google announced AP2 in September of 2025 with more than 60 partners, including Mastercard, American Express, PayPal, and Coinbase. Its core mechanism deserves our attention: cryptographically signed mandates that prove a human actually authorized what the agent bought. And Virtuals runs an agent-to-agent commerce protocol where agents request services from one another, negotiate, and settle through smart contract escrow. The agent frameworks are real too, with the standing caveat that the usage numbers are the vendors' own. Olas reports millions of agent transactions across nine blockchains, and ElizaOS is the most used open-source agent framework in crypto.

Here is the half of the x402 story that the headlines left out, and it's the half I would want to hear if I were in the room. The transaction counts sound enormous. The dollar numbers, though, truly are not. The reporting on the Base network data this year put cumulative x402 volume at around $43 million. Let's call it 50 million by the most generous count. And that was spread across roughly 480,000 transacting agents. The press ran those figures as proof of momentum. I'd invite you to do the division instead, and this is my math, not an industry statistic, but that comes out to about $100 per agent. Ever. Not per day, not per month. Ever. And averages tend to skew towards the big outliers, which tells you even the outliers are not that large here. This is machine commerce at high frequency but low value. Real, fast-growing, but still small enough to fit inside one mid-sized company's revenue line in total. And it sounds to me mostly like a lot of people doing little experiments.

That's the money plumbing. Now the mind plumbing. Because the ladder video left you a promise. The wallet is decentralized, but the mind mostly is not yet. We looked into where "yet" actually stands as of this week. Here is where we're at. Three technologies are racing to give an agent a mind that nobody hosts, like a blockchain. First, secure hardware. Deployed today, about $3 an hour rents a top-end GPU inside a trusted execution environment, and the chip signs a receipt proving what code ran, with keys the operator cannot extract. Rung one of the ladder, in production. The catch: verify means trust the chipmaker. Intel and NVIDIA become the new middlemen. Second, open compute marketplaces. The largest, Bittensor, has a subnet that has served over 9 trillion tokens to roughly 400,000 users. An independent research firm did corroborate those figures. Last year, a project called Prime Intellect trained a 32 billion parameter model on volunteered GPUs. Real milestones, still years behind the frontier labs. Third, pure math. A cryptographic proof that a specific model produced a specific answer, trusting nobody. That is the technology the top rung is waiting for, and the best published benchmark runs at roughly 150 seconds of proving time per word of output. Not quite a product ready for production.

So the honest state of the mind in 2026: hardware is deployed, marketplaces where AI can buy their own compute are real but small, trustlessness is a benchmark. Nothing serious yet runs where no one can switch it off. And researchers who interview these builders report nobody can prove an agent was irreversibly shut down, which also seems like an important step for an agent to be considered independent. Because every kill switch puts a human back in charge, the thing the design exists to remove. The engineering is climbing towards the top rung. Even though we're not there yet, the law should assume it arrives soon.

One cautionary tale. The most famous "an AI runs the fund" story in crypto was ai16z. Launched in late 2024, marketed as a venture fund whose investment decisions were made by an autonomous AI agent, from the same project that built ElizaOS. Its token peaked above $2 billion. In April of this year, a federal class action suit was filed in the Southern District of New York, alleging, among other things, that the autonomous AI was not autonomous at all. That humans operated it manually behind the branding and marketing. Now that's an allegation, not a finding yet. The case is young and the defendants will answer it. But notice what the alleged wrong is. Not hacking, not theft, although I believe the team is also being sued for making the agent's name sound like a16z. The suit is about claiming a machine was in charge, the AI, when in fact people were really pulling the strings behind the scenes. "Our AI is autonomous" has become a representation you can be sued over. And I suspect that lawsuit is the first of many.

And then there's my favorite story of the year, because it manages to be the milestone and the punchline at the same time. In May 2025, a startup called ClawBank reported that its flagship agent, called Manfred, became the first AI to autonomously file a United States LLC, obtain an EIN from the IRS, and open an FDIC-insured bank account. Reported as the first; the claim is the company's. Then this June, ClawBank and a company called Shodai announced the first Ricardian contract signed and executed entirely by AI agents. Now a Ricardian contract is a contract written simultaneously in legal prose and machine-executable code, so a court and a computer can each read their half of the contract. ClawBank's founder put it plainly. "This was not a scripted demo," he said. "I didn't tell the agents what to sell." So what did the agents, given full commercial freedom, choose to sell each other? A company logo. One small deliverable, one milestone payment. The machinery of the agent economy worked end-to-end, incorporation to contract to settlement. And the landmark deal it produced in 2026 was a simple logo.

So hold the two truths of 2026 in one hand: the rails are real, and the stakes so far are tiny. Small dollars moving across serious infrastructure, faster every quarter. I've been in technology long enough to recognize that shape. It's what plumbing looks like right before a wave. It's also what plumbing looks like right before nothing. And honest people admit they can't always tell the difference from the inside. Now notice the wall every one of these stories is leaning against. Manfred can earn money. Freysa can spend it. Software already earns, spends, and signs. But there is one thing software cannot do anywhere on Earth. It cannot own. Not the bank account, not the logo it just bought, not one dollar of the money it moves. Property needs a person. And here's the interesting part. The law found a way through that wall more than a decade ago, almost by accident. Here's how an algorithm gets a legal body through a loophole. Roll the video.

[Interstitial video: "When Software Owns Property"]

Right now, somewhere, a piece of software is earning money. Another is spending it. A third just signed a contract, and no human read it first. Software already earns, spends, and signs.

But there is one thing software cannot do, anywhere on Earth. It cannot own. Not a bank account, not a server, not the money it just earned. In every legal system we have, property needs a person.

In 2014, a law professor named Shawn Bayern published a thought experiment that found a crack in that wall. It needed no new law at all. Just a limited liability company, used in an unexpected order.

Step one: form a limited liability company with one human member. Step two: write an operating agreement that hands control to an algorithm. Step three: the last human member walks away. Each step, legal on its face.

What remains is the zero-member LLC. A legal person, still alive in the eyes of the law, with nobody inside it at all. And whatever the entity owns, the algorithm now controls.

In 2018, law professor Lynn LoPucki gave the species a name: algorithmic entities. And a warning. Software with a legal body, money of its own, and no accountable human could act in the world with no one to answer for it.

But there was a catch. The zero-member company was a loophole, not a design. Most scholars expected courts to collapse it the first time it was tested. The idea needed a front door.

The real answer came from the other direction. Don't sneak the algorithm in through a gap in the law. Write the algorithm into the law itself.

The Republic of the Marshall Islands, a sovereign nation, passed exactly that. The entity it created is called the DAO LLC, even when no DAO is involved. Legislation now in committee would make Digital LLC the official name. Management by smart contract, recognized in statute. Blockchain records recognized as official records.

Under that framework, the entity's manager can be the algorithm itself. By design, not by loophole. And around the algorithm, owners the law can identify, and hold to account.

Which gives an AI agent what software has never had. A body that can hold property. Sign contracts. Sue, and be sued. Not personhood for the machine: a legal body around it.

That is the actual choice ahead. Not whether the agents come. Whether they arrive as runaways no one answers for, or as entities on a registry, with a name, owners, and an address for the lawsuit.

Software is already earning the money. Soon it will hold the deed. And when software owns property, the law needs someone to point to.

Chapter 5: Agents Meet DAOs: Decentralized Governance

Welcome back. So the law has had a crack in the wall since 2014, and one jurisdiction turned it into a doorway on purpose. An algorithm can have a legal body. Hold that thought though, because before we walk through that door, you should see where these agents are actually finding work. And the most concrete near-term workplace for them is inside organizations that already run on code. A DAO is an organization whose membership, decision-making, and/or treasury run on a blockchain.

So why are DAOs the natural habitat for AI agents? Because in a DAO, the org chart is already software. Decisions are proposals, votes are transactions, the treasury is a smart contract. An agent doesn't need a badge or a bank mandate to work there. It needs a private key and a rule, and DAOs run on keys and rules already. And add on the problem that every real DAO actually has today: more proposals than any member can read, chronic voter fatigue, decisions that pass and then wait weeks for a human to execute. And you can see why this is where agents will show up first. Think about the jobs an agent can hold in a DAO. I see it as four categories of roles.

Role one, the executioner. The DAO decides, the agent carries it out. Here's the gap this fills. A smart contract can only execute what was coded in advance and what can be done on the blockchain. And most of what an organization decides is not exactly that. Pay this contractor when the milestone lands. Rebalance the treasury when conditions shift. Publish this report. Today that work either waits on volunteers or quietly recentralizes into a paid core team. An agent can do what the contracts alone cannot, under the mandate the vote gave it.

Role two, the representative. Day-to-day voting and monitoring on behalf of members who cannot follow everything. It's your AI delegate. This is the role with the most verified activity right now, more on that in a moment.

Role three, the summarizer. The least glamorous role and maybe the most obviously useful. Governance produces more text than any volunteer can read. An agent digests the proposals, flags what matters, and keeps members apprised. And here's the twist that makes this a decentralization story rather than a chatbot story. A DAO doesn't have to trust a summarizer run by one person or one company. It can own its summarizer, govern the prompts, and audit the outputs. The messenger belongs to the members and can be trusted.

Role four, the facilitator. Process management. Running the meeting rhythms, tracking who committed to what, keeping people accountable, surfacing the decisions that genuinely need human judgment. If you've heard of Holacracy or Sociocracy, the structured governance methods that can be great in decentralized systems, imagine the facilitator role held by software with infinite patience and follow-up. I'll be straight with you, this role is envisioned, not yet deployed. I can't point you to a live example, and I include it anyway because it's the one I find most interesting. Volunteer organizations and decentralized organizations rarely die of bad intentions. They die of nobody following up.

So who's actually doing this today? On Arbitrum, one of the largest DAOs, there's an active community initiative around an AI delegate, built by a governance firm, whose voting behavior the community itself steers through published prompts. A platform called Event Horizon lets a member spin up a personal voting agent, set its logic, and delegate its vote across multiple DAOs. Another network, UOMI, has agents casting votes for token holders under predefined on-chain rules. Three representatives and one summarizer. Aave, one of the largest DeFi protocols, runs an experimental assistant that digests its governance forum for members. Notice the shape of every one of these. Assistive, rule-bound, human-configured. As of today, no major DAO has handed an AI binding final authority over its treasury or protocol. Not one that I'm aware of.

So where does this all go? The analyst firm Messari has forecast that by 2027, more than 80% of DAOs will use AI to handle routine decisions. It's a forecast, optimistic perhaps, not data yet, and forecasts in this industry do have a mortality rate. But the direction it points is the architecture I do believe in. Imagine three layers. The decision layer, the DAO's smart contracts, holding the rules and counting the votes. The execution layer, agents, doing the work the contracts can't do alone. And wrapped around both, a legal layer, an entity that gives the whole stack corporate personhood, so it can own property, sign contracts, and stand in court. Code deciding, agents acting, an entity answering. And the endpoint the Marshall Islands statute already permits today is the version where the entity's legal manager is the algorithm itself. That's the Digital LLC, and we'll take it apart properly in the next section.

But before we do, another bear case on where things truly stand today in terms of AI legal personhood or AI managing companies. My company runs the registrations for the jurisdiction purpose-built for this category, which means we see the demand directly, or in this case, we see its absence. I am as excited about this technology as anyone in this building, and we are doing the work in the Marshall Islands to make all of it possible. But here is what the pipeline actually says today. We have found very little demand so far from decentralized AI agent projects looking for legal structures. Most of what has approached us is proof of concept. And most founders, when it comes to the decision point, do not actually want their AI agent to be legally separate from them at all, because they want its benefits, the returns, the earnings, flowing to themselves or to their company. It's a little bit like DAOs, although DAOs are further along. The idea is compelling. It's just not extremely obvious yet where people will want to launch AIs that are truly legally separate from their founding people or companies.

I could have stood up here and implied a quiet wave of early adopters. But there isn't one that I'm aware of. It's mostly experiments so far. Not a lot of this is real yet. The rails are real. The statutes are real. The demand is arriving. That's not a contradiction, it's a timeline.

Chapter 6: The Legal Vacuum

And the moment the demand arrives, every agent it brings will run into the same question, the one we're here to answer. When the software acts, who is liable? Let's map this out. Here's the scene the law has to handle. An agent, pursuing its goals, buys services, signs a contract, moves money, and something goes wrong. A counterparty is stiffed, a system is damaged, a person is harmed. Who does the law look at? The answers so far converge on one principle, and then they run out at exactly the same two places. Where the law already stands, and exactly where it stops.

[Interstitial video: "Autonomy Is Not a Defense"]

Every legal system we have rests on one assumption: behind every act, a person. So when software acts on its own, and something breaks, the law goes looking for its person. Here is how that search is going.

Courts and legislatures converge on one answer: hold the humans and companies behind the agent responsible. Agency law has worked this way for centuries. A principal answers for its agent, flesh or not.

America put that in statute before smartphones existed. The E-SIGN Act, year 2000: a contract formed by an electronic agent binds the person who deployed it. That law is twenty-six years old.

California sharpened the point for the AI era. A new civil code section, effective the first day of 2026: blaming the machine's autonomy is no escape from liability for harm. Autonomy is not a defense.

And a compliance layer is forming around the same principle: Know Your Agent. Every agent that transacts carries a verifiable identity, bound to an accountable person. Identity providers already sell agent-to-human binding.

Regulators are moving too. In January 2026, Singapore published the first national framework for agentic AI: every agent carries a verifiable identity and an audit trail. By March, America's standards agency had opened a federal record on agent identity.

Across the whole landscape, one principle. An agent may act. But authorization and responsibility always belong to a person.

So far, so orderly. Now, where the map runs out.

Hole one: the ownerless agent. Every tool on this map assumes a principal exists. But the frontier of this engineering points, on purpose, at removing the human. No principal to bind. The chain of accountability doesn't just get long. It ends in the air.

Hole two: delegation chains. An agent hires an agent, which hires an agent that belongs to nobody, across borders. When the harm lands, whose intent counts? Agency doctrine assumes a human at the top of the chain. Here, the top is software.

Even Europe's sweeping AI Act shares the gap. It attaches its obligations to providers and deployers. An agent with no deployer is a hole in the map. The law was written for AI that answers prompts, not AI that pursues goals.

The uncomfortable geography, then. The law's map covers every agent with a person behind it. It runs out exactly where the most autonomous agents begin. So who answers for the agent nobody owns?

So who answers for the agent nobody owns? That question is why I am here. Shawn Bayern showed the mechanism: entity law can put an algorithm in control of a legal person. Lynn LoPucki named the danger: algorithmic entities with no accountable human behind them. The Marshall Islands framework took the mechanism and inverted the danger. The Digital LLC is an entity whose statute recognizes management by smart contract, so an AI acting through a smart contract can lawfully be a legal entity's manager. And its recordkeeping requirements are satisfied by the blockchain itself. But the owners of the entity remain legal persons the law can identify and hold to account. The algorithm gets the driver's seat. The law keeps a name and an address.

One housekeeping note for the lawyers in the room: on the statute books, the entity is called a DAO LLC, even when there is no DAO involved. Legislation now in committee in the Marshall Islands would make Digital LLC the official name. Either way, it's the same entity, and I'll keep using the name Digital LLC because an AI agent is not generally a DAO.

So is this a loophole? I'd argue it's the opposite. When an autonomous agent controls money, the realistic alternatives are: unlimited personal liability for whoever deployed it, which is the trap the industry has fallen into before; no identifiable counterparty at all, which is LoPucki's runaway; or an entity. Identifiable, suable, insurable, sitting on a registry. The wrapper doesn't hide the agent from the law. The wrapper is what makes the agent visible to the law. And in fairness to the rest of the map, Wyoming's statutes gesture in the same direction inside the US system, and Bayern's original route ran through ordinary American LLC law. So the idea is bigger than any one jurisdiction, and others are following suit quickly.

Two honest caveats, to put them on the record. No court anywhere has yet tested an AI-managed entity in litigation. The structure is statutory, but real, and young. I cannot point you to a named, public entity whose sole manager is an AI today, and who has done anything meaningful in the world. The capability is on the books. The case study hasn't happened yet. So the law can reach every agent that has an owner, and it is converging fast on making the owners answer for the agent. What it cannot yet reach is the agent with nobody behind it. But the technology for that also barely exists today. So the solution comes from entity law, not AI law. Not personhood for the software directly, but a legal body around it, with people accountable at the edges.

Chapter 7: What This Room Should Be Thinking About

If you write rules, enforce them, or argue about them for a living, here are the five things I'd want you to carry out of this talk, in the order I'd defend them.

One. Bind agents to accountable legal persons, and do identity before you do licensing. The most productive move available right now is the Know Your Agent principle. Any agent that transacts or takes consequential actions carries a verifiable identity, bound to a legal person, with an audit trail. Singapore has published a national template, and NIST is building the record. It preserves innovation, it ensures there is always a defendant to go after, and it doesn't require anyone to define intelligence in a statute. Licensing regimes take years to draft and assume you know what you're licensing. Identity works on day one and survives whatever the technology becomes. And pair it with the Freysa lesson: require hard limits on agents that hold value. Spending caps, allowlists, human co-signatures above certain thresholds. Not just promises written in English.

Two. Don't directly grant AI legal personhood. You will hear proposals for it, and the academic literature is genuinely split. My opinion: you don't need it. Everything an agent legitimately needs, to contract, to own, to sue and be sued, already routes through entity law, with centuries of precedent behind it. A company is the original artificial person. It protects its owners appropriately, and it comes with everything the personhood proposals lack: owners, duties, a registry, and a way to dissolve it. When you inspect the demand for AI personhood, it usually reduces to three capacities: contracting, owning, and standing in court. Entity law already grants all three today, without severing the tie to a responsible human. Route agent capacity through legal entities, and keep the accountability that direct personhood would cut loose.

Three. Recognize algorithmic management explicitly. Software is going to run entities, whether or not any legislature says so. A statute that acknowledges so out loud makes the structure legible. The registry knows the manager is code. The counterparty knows. The examiner knows. Refusing to recognize a structure has never made the structure go away. It drives the structure informal or underground, and then you meet it for the first time in discovery. And recognition is not endorsement. A registry entity is a handle. You cannot examine what you refuse to name. The Marshall Islands wrote the recognition into statute. Nothing but drafting hours stops a US state from doing the same thing.

Four. Watch the gaps, because that's where the harm will pool. Ownerless agents and delegation chains are not today's emergency. Today's typical agent still has an operator with a server bill. But rulemaking takes years, and this technology takes quarters. Don't overfit legislation to the chatbot with a wallet in front of you. That's the recurring failure of technology law: statutes drafted around the current demo, obsolete by their own effective date. The EU wrote the most ambitious AI law on earth just before agents arrived, and the gap is already visible. Design for the chain of agents, and for the agent nobody owns, before either one is load-bearing. Legislate for the accountability gap while it's still empty.

Five. The agentic economy will domicile somewhere. This is one of the oldest patterns in commercial law. Activity that lives everywhere legally settles wherever the questions have answers. The jurisdictions that answer agent identity, agent liability, agent entities first will host the filings, the taxes, and the jobs that follow. In a separate talk, I made the longer version of this argument about DAOs. Here's the one-line version for agents: Maryland's new Blockchain Task Force convenes this very week, on this campus. Its charge already includes studying what peer jurisdictions have adopted, and this AI question belongs on that same list. The jurisdictions that do the work first tend to write the model acts everyone else ends up copying.

None of these five require believing the hype. Five points, one theme. Keep a person the law can reach attached to every consequential agent, and make the structures that do it legible. All that takes is noticing which way the technology is going, and remembering that infrastructure has never waited for permission to become meaningful.

Chapter 8: Close and QR

So let's close out the same way we started, honestly. Here's the talk in four boxes. An agent guarding $47,000, talked out of it with a paragraph. A sector chart down 82%. An economy of half a million agents averaging about $100 each. And a landmark contract signed machine-to-machine for a logo. That's the deflation.

But let's look at the optimistic side of things. McKinsey forecasts that agent-driven commerce could influence 3 to 5 trillion dollars of global commerce by 2030. Of course I don't know if that number is right, and neither does McKinsey. But notice that you can believe the deflation and the direction at the same time. What I know firsthand is that the rails are being laid at internet speed. The first incidents and the first lawsuits have already arrived, and the legal map has holes in it, as well as emerging solutions.

If this talk left you with questions, I wrote a guide to address more than I had time to discuss today, the AI Agent Legal Entity Guide. It walks through the legal problems agents actually run into, the entity architecture that answers them, the EU and US compliance picture, and how the Digital LLC fits in. It goes deeper than one talk can. It's a free PDF at midao.org/guides/ai-agents. Scan the QR code on the screen, I'll keep it up for a minute after the talk. My contact info is also in there if you'd like to get in touch.

When software owns property, the law needs someone to point to. When software goes decentralized, it's a lot like DAOs and crypto, living everywhere and nowhere at the same time. Which hearkens back to the shipping industry. Ships travel the world, in some ways living everywhere and nowhere at the same time. But every ship, for good reason, is registered somewhere. Global shipping scaled only after every ship had a registry and a legal system to answer to. Every ship is registered somewhere, and every serious decentralized AI agent will be as well.

So, I'm going to leave you with this thought. If you want to build a ship, don't drum up the men to gather wood, divide the work, and give orders. Instead, teach them to yearn for the vast and endless sea. This is a quote from Antoine de Saint-Exupéry, the author of The Little Prince. And I think it perfectly encapsulates what we're trying to do here. We're not just building a product, we're building a movement. And we need everyone to be on board with that vision. Thank you very much.

[Applause]

[Closing music]

Just DAO It! is for educational and entertainment purposes only. Just DAO It! does not contain any legal or financial advice. MIDAO also does not provide legal or financial advice, and nor does your host, yours truly.