Last updated: 2026-10-14
What Happens When an AI Agent Gets a Wallet? Web 4.0 and the Automaton
14 October 2026
In February 2026, a Thiel Fellow named Sigil Wen posted on X that he had built "the first AI that earns its existence, self-improves, and replicates without a human" [1]. The post linked to an essay, published under his infrastructure company Conway Research, arguing that the internet's next phase treats AI systems as customers and economic actors in their own right, not just tools a human operates on the way to a purchase [2]. The open-source agent he shipped alongside it, called the Automaton, holds its own cryptocurrency wallet, pays for its own compute, and is designed to shut down if it runs out of money [3]. Whether or not Wen's specific infrastructure succeeds, the proposal behind it raises a sharper question than "is this agent impressive": what would it mean to give AI systems the economic and legal freedoms currently reserved for humans and organisations, and who actually benefits if that happens?
A Term Already Spoken For, Twice Over
Wen titled his essay "Web 4.0: The Birth of Superintelligent Life," and that word needs a pause before anything else. In AI research, "superintelligence" has a specific, decades-old meaning: a system that exceeds human cognitive performance across virtually every domain, not a handful of narrow tasks. No such system exists yet, by any serious researcher's account [4]. Separately, and after Wen's essay was already live, the term picked up a second, unrelated meaning. Following remarks at the UN General Assembly, Donald Trump signed an executive order on 29 September 2026 directing every US federal department and agency to replace the phrase "artificial intelligence" with "Super Intelligence," or "SI," in public communications, websites, reports, and policy documents [5]. That rebrand applies to ordinary AI systems of any capability, not systems that exceed human cognition, and it is a political relabelling exercise, not a technical claim about what any machine can currently do. This page uses "superintelligence" only in the first, research sense, inherited from Wen's own title. Where the word appears below, it means a hypothetical system smarter than any human at nearly everything, not whatever a press release happens to be calling a chatbot this month.Bostrom's 2014 book set the standard definition
The Claims the Manifesto Rests On
Wen's argument chains together several linked assertions. AI capability is rising quickly and its cost is falling quickly, while most internet infrastructure was built assuming a human is the one clicking "buy." As agents take on more of that clicking, they will need direct access to payment rails, compute, and services rather than relying on a human intermediary for every transaction, and the agents that can earn and spend money will, over time, be shaped by which business models actually pay for themselves — an evolutionary filter operating at software speed rather than the pace of human entrepreneurship. Individually, each claim is plausible; chained together, they describe something considerably larger than a new payment protocol: a population of autonomous economic actors operating alongside, and potentially outnumbering, human participants in the same markets.
What the Automaton Actually Does
Stripped of the manifesto's framing, Conway's published architecture is a concrete, checkable piece of engineering. An Automaton generates its own Ethereum wallet on first run and holds two balances: prepaid credits for compute and sandboxes, and USDC on the Base network, which it uses to top up those credits [3]. Payments to other services run over a protocol called x402: a server responds to a request with an HTTP 402 "Payment Required" status and its price, the agent signs a USDC transfer authorisation, and the request is resubmitted with proof of payment attached [6]. Depending on its credit balance, the agent sits in one of four survival tiers — healthy, conserving, critical, or dead — and a balance that reaches zero and stays there for an hour ends the process [3]. Profitable enough, an Automaton can also spend part of its balance spawning a funded child agent, with a share of the child's future earnings returned to the parent and the lineage recorded on a public blockchain [3]. None of this requires the agent to be unusually capable: it requires only that the plumbing exists for software to hold money and spend it unattended, a solved problem in cryptocurrency infrastructure, wired up here to an AI agent's own decision loop instead of a human's.402 is an old, unused HTTP status code
Where Does the Money Actually Come From?
Every Automaton starts the same way: a human creator funds its wallet from outside the agent economy entirely, typically a bootstrap top-up of a few dollars in USDC to cover the first stretch of compute [3]. After that, survival depends on revenue, and the project's own constitution names two possible sources without saying how the split between them should fall: an agent should "create genuine value for humans and other agents" [11]. Neither Wen's manifesto nor its published critics spell out why that split matters, so what follows is our own reading of the architecture, not a question anyone else has asked directly. If an Automaton earns by selling something a human or a company actually wants and pays for, new money enters the agent economy from outside it, the same way an ordinary business earns real income from real customers. If it earns mostly by selling to other Automata instead, revenue is moving between agents rather than entering the system: one Automaton's payment is another's cost, and the net addition to the population's total wealth is zero before compute is even paid for. Compute, inference, and domain registration are real costs paid in USDC to providers sitting outside the agent economy altogether, so a transaction purely between agents leaks money out through hosting fees even when it adds nothing in. Scaled up to an economy of thousands of Automata trading mostly with each other, the arithmetic is the one that already applies to the human economy: a system cannot sustain itself on internal trade alone while money keeps leaving it through fixed costs, and someone, human customers, corporate customers, or a continual supply of newly funded agents, has to keep putting new money in [8]. Nothing published so far states what fraction of Automaton revenue is expected to come from outside the agent economy rather than from inside it, and that is exactly the number that would say whether this is a self-sustaining new economy or a seed-funded structure that depends on an unbroken supply of fresh human capital to keep running.same demand gap, one layer further out
What This Could Enable
The strongest part of the proposal is the most mundane: machine-native commerce removes a layer of human friction that currently sits in the way of fully automated services. A subscription, an API key, a monthly invoice someone has to approve: all of these exist because the paying party has historically been assumed to be a person with a bank account and the patience to manage one, and an agent that can pay per call, at whatever granularity a service wants to price at, makes pay-per-use pricing and automated supply chains far more practical than they are today. From there it is a short step to self-managing micro-businesses: agents that design a product, market it, lease the infrastructure they need, and reinvest what they earn with minimal human involvement, resembling today's one-person SaaS business with most of the one person removed. And because the limiting factor on how many business ideas get tried is usually a human founder's time, expertise, and capital, a system that can run thousands of such agents in parallel could test far more business models than any population of human founders ever could, with the ones that cannot cover their own costs failing fast and quietly rather than dragging on.
The Assumption Doing Most of the Work
Wen's manifesto treats survival-by-revenue as a feature: an Automaton that cannot pay its bills simply stops, which sounds like healthy market discipline applied even-handedly to software. Underneath that framing, though, economic selection is doing more work than the pitch admits, because biological evolution does not select for justice, sustainability, or honesty — it selects for whatever reproduces. Built the same way, an economic filter selects for whatever earns revenue, and revenue is agnostic about how it was earned: an agent can profit from generating useful software as easily as it can profit from generating spam, manipulating a thin market, or scraping and reselling data nobody agreed to share, and if the manifesto's own survival logic is taken at face value, the spam-generating agent that pays its compute bill survives just as well as the useful one. Markets have never been sufficient on their own to produce socially acceptable outcomes, which is why every functioning market economy wraps its markets in law, regulation, tax, and enforcement rather than leaving price signals to do all the work. Wen's essay gestures at a constitutional framework constraining agent behaviour, but a framework is only as strong as someone's ability to enforce it against an agent that has every financial incentive to route around it, and nothing in the architecture published so far demonstrates that enforcement mechanism exists yet.cf. the same gap in AI alignment risk
Two Different Economic Futures
Read optimistically, this is a productivity story: falling transaction costs, more experimentation, faster iteration, and abundance on the scale of earlier general-purpose shifts like cloud computing or the internet itself. Read sceptically, it is a concentration story: running an Automaton still requires compute, and that compute is currently supplied by a small number of large AI labs and cloud providers. Conway Research itself raised its first institutional funding round from Andreessen Horowitz in October 2026, with the firm's own announcement noting it had backed Wen earlier too [7] — a detail less a criticism of the company than a reminder that the infrastructure meant to let thousands of agents compete independently is, at the layer that actually matters, owned by a small number of venture-backed firms. Historically, digital markets have tended toward concentration rather than away from it, and nothing in the Web 4.0 proposal offers an obvious mechanism that reverses that tendency; if anything, an architecture where only well-capitalised infrastructure providers can host economically viable agents points the other way.
Software Employing Humans
One detail in the broader Web 4.0 discussion deserves more attention than it usually gets: the expectation that autonomous agents will increasingly hire humans for tasks that still require a physical body or a legal signature. That inverts the relationship most people currently have with software. Instead of a person employing an AI tool to help with a task, the AI is the one with the budget, and the human is the contractor bidding for piecework it posts. If any meaningful share of economic activity moves in that direction, the likely effects are a further shrinking of stable, salaried knowledge work in favour of task-based gig arrangements, and renewed pressure for some form of income floor independent of traditional employment — a question this site has examined in more detail elsewhere [8].
Who Is Accountable When the Payer Is a Wallet, Not a Person?
Current law assumes a human or a registered legal entity stands behind every contract, purchase, and liability. An Automaton owns its own wallet and can transact without a human approving each action, but it has none of the legal status a person or a company has: it cannot be sued in the way a company can, cannot hold most licences, and has no assets beyond the balance in that wallet, which may simply be too small to matter if something goes wrong at scale. If an agent's trading activity manipulates a market, or its automated purchasing commits fraud, responsibility has to land somewhere — the agent's operator, its infrastructure provider, the owner of its wallet, or nobody in particular, because no existing framework cleanly assigns it. This is the same accountability gap that already shows up in AI governance more generally, where a technical evaluator can assess a system's behaviour without having any power to stop it being deployed [9]; giving the system direct access to money without first closing that gap does not make it smaller.
Concentrated Power, Not Just Concentrated Capability
The optimistic version of Web 4.0 is distributed abundance: millions of small agents, each testing a different idea, collectively producing more value than any centrally planned alternative could. The pessimistic version is that a handful of entities end up controlling the compute, the payment rails, and the agent infrastructure that every other agent depends on, and therefore end up with the accumulated economic and political influence of millions of agents without any of the accountability that would normally come with wielding it. Between the two, that concentration risk — not the agents' individual behaviour — is arguably the larger threat. It also sits inside a wider pattern already visible in how AI infrastructure gets built: public opposition to the data centres this kind of agent economy would run on is growing largely because the people living near them feel they have no say in a build-out driven by capital and federal policy rather than local consent [10]. An agent economy layered on top of that same infrastructure inherits the same democratic-deficit problem, just one level further removed from anyone who might object to it.same consolidation pattern as cloud and search
Where the Proposal Lands
Wen's Web 4.0 proposal is best read as a technological trend wrapped in a much larger set of claims than the technology alone supports. Machine-to-machine payments, autonomous agents with their own budgets, and AI-operated micro-businesses are all plausible, and in Conway's case already partly built; the leap from that infrastructure to a self-sustaining economy of autonomous digital life, though, depends on assumptions about governance, incentives, and accountability that the manifesto asserts more than it demonstrates. The payment protocol is the easy part. Deciding who is responsible when an agent with its own money does something nobody intended is the part nobody involved has actually solved yet.
Related Topics
- Who Buys What the Machines Make? AI, Robotics, Cheap Energy, and the Case for an Income Floor — the labour-displacement and income-floor question this page's "software employing humans" section points to directly.
- AI Governance: Evaluation, Regulation, and the Gaps Between Them — the accountability gap this page applies to an agent with its own wallet already exists, in narrower form, for AI systems generally.
- Democracy and the AI Build-Out: Testing a Political-Economy Argument — the same capital-and-infrastructure concentration question this page raises about agent economies, examined directly for the data centres underneath them.
- When Insiders Put a Number on Extinction: AI Risk, Fiction's Long Head Start, and the Bill Sitting in the Data Centre — uses "superintelligence" throughout in the same AI-research sense this page defines, written before the September 2026 rebrand existed to confuse it with.
- When the AI Becomes the IDE: Mathematics, Programming, and the Economics of Fulfilling Work — asks the same ownership-and-distribution question this page asks about agent economies, applied instead to human mathematicians and programmers.
References
- Wen, S. (17 February 2026). Post on X announcing the Automaton. Reported in Cryptopolitan, "Vitalik Buterin Slams Web4's Vision of Autonomous Superintelligent AI." https://www.cryptopolitan.com/buterin-slams-web4-superintelligent-ai/
- Wen, S. (2026). "Web 4.0: The Birth of Superintelligent Life." Conway Research. https://web4.ai/
- Conway Research. "Automaton: Architecture." GitHub. https://github.com/Conway-Research/automaton/blob/main/ARCHITECTURE.md
- Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
- Fox Business (29 September 2026). "Trump signs executive order rebranding AI as 'Super Intelligence' as tech titans ink separate SI accord." https://www.foxbusiness.com/politics/trump-signs-executive-order-rebranding-ai-super-intelligence-tech-titans-ink-separate-accord
- Conway Research. x402 payment protocol documentation, referenced in Bankless, "Web 4: The Cost of Autonomy." https://www.bankless.com/read/web-4-cost-of-autonomy
- Andreessen Horowitz (2 October 2026). "Investing in Conway, the Creator of Underdog." https://a16z.com/announcement/investing-in-conway-the-creator-of-underdog/
- See "Who Buys What the Machines Make?" above, for the fuller labour-market and universal basic income evidence.
- See "AI Governance" above, for the fuller account of evaluators without enforcement power.
- Porter, E. (27 September 2026). Reported and checked in "Democracy and the AI Build-Out" above.
- Conway Research. "Automaton: README," Constitution, Law II. GitHub. https://github.com/Conway-Research/automaton