What Accelerando got right about the autonomous AI company
The Cover for Accelerando by Charles Stross
I. A man walks through Amsterdam with a company in his glasses
Charles Stross opens Accelerando with a man crossing a plaza in front of Amsterdam’s Centraal Station, sunlight coming off the canal, his eyeballs powered up. Manfred Macx is a “venture altruist.” He generates business ideas at an absurd rate, patents them, signs the patents over to an open-source trust, and gives the value away. He owns almost nothing. He carries no meaningful money. He lives extremely well anyway, because he has made so many other people rich that they keep him supplied out of gratitude and self-interest. His cognition is distributed across a wearable rig and a scattering of remote machines — an exocortex — and when it is stolen from him later in the book, he loses not just his files but a substantial portion of his self.
The chapter, “Lobsters,” first appeared in Asimov’s Science Fiction in June 2001. Eight more novelettes followed through 2004, and Ace published the collected novel in July 2005. Stross released it free under a Creative Commons license, which is why the whole thing is still a click away. It won the Locus Award for best science fiction novel in 2006.
Reading it in 2026, the striking thing is not the hardware. The smart glasses are here already. The distributed cognitive prosthesis is here and is called a phone. What holds up — what has gotten sharper rather than duller with twenty years of hindsight — is Stross’s argument about the shape of the economic entity. He saw that the interesting frontier was not machines that think. It was machines that own.
Everything downstream in the novel follows from that. Amber Macx, Manfred’s daughter, escapes her mother’s custody in the chapter “Halo” by indenturing herself to a Yemeni shell corporation that she herself controls, then signing aboard a mining ship in the Jovian system. She does not hack a computer. She hacks a jurisdiction. Later she claims sovereignty over a privately owned asteroid, converting a property title into a state. When the crew finally reaches the alien router beyond the solar system, the entity waiting for them is a thing Stross’s own character list describes as a corporation, an economy, and a 419 scam simultaneously — the Slug, an intelligence whose entire method of predation is contract law. And at the far end of the book, the posthuman intelligences of the inner system, the Vile Offspring, take the planets apart to build a Matrioshka brain, a nested computational shell around the sun.
Stross has been blunt about what the arc means. In a 2013 note on the novel’s construction, he described a book that looks like techno-optimism while horrible things happen in the background: capitalism eats everything, and the logic of competition eventually pushes past the point where merely human entities can compete. We become, in his phrase, a slow and tasty resource. Something like the dodo.
That is the cautionary half. But there is another half, and it is the reason the book keeps coming back to people who build things: Manfred is not a victim of this system, and he is not a predator inside it either. He is running a third strategy. He assembles ventures at speed, extracts almost none of the surplus, and pushes the value outward into the commons, where it compounds in ways he does not control and cannot capture. The most agentic character in the novel is also the least acquisitive one. That is a design choice, and it is available to us.
What follows is an argument that the machinery for Manfred’s kind of company — the ad hoc entity that assembles itself around an opportunity, operates continuously, and disbands cleanly — now largely exists. The legal substrate exists. The payment substrate exists. The identity substrate is being poured. The reasoning substrate is improving on a measurable curve. What does not yet exist is the judgment to run one unsupervised, and the harder thing: a reason for one to exist that is worth the trouble.
II. The oldest artificial intelligence
In December 2017, at the 34th Chaos Communication Congress in Leipzig, Stross gave a keynote called “Dude, you broke the future!” He set aside the singularity as an explanatory model and offered a substitute. Looking at the last two to four centuries, he said, the glaring deviation from the previous three thousand centuries is the invention of artificial intelligence — which happened no earlier than 1553 and no later than 1844. He was talking, of course, about “the very old, very slow AIs we call corporations.” Ted Chiang had made a version of the same observation around the same time; so had others. But Stross pushed it further than a bon mot, because he had already spent a novel working out the consequences.
This is how the argument goes: A corporation is an artificial person. It is created by filing. It has goals that are not the goals of any human inside it. It has perpetual succession — it does not die when its founders do. It can own property, sign contracts, borrow, sue, and be sued. Its component humans are, functionally, interchangeable and replaceable, retained only so long as they perform their assigned role. William Kyd’s 1793 treatise on corporations described exactly this: a collection of individuals united into one body under an artificial form, with perpetual succession, vested by law with the capacity to act as an individual.
Nothing in that description requires the components to be human.
This is the insight that Accelerando dramatizes and that our current moment makes urgent. We have spent three years asking whether language models can reason, plan, and act. It is the wrong question to start with, or at least the wrong one to start with if you want to build. The prior question is: what does a piece of software need in order to hold economic agency? And the answer, historically, is not a body, a bank relationship, or a legal theory of machine consciousness. It is a wrapper. A container recognized by other institutions as a locus of rights and obligations.
Manfred’s genius, in the novel, is that he treats corporate form as a programmable primitive — something you instantiate, use, and discard, the way a developer instantiates an object. Amber’s escape works the same way. Stross understood that once you can spin up legal persons cheaply, and once software can direct their actions, the distinction between “software that runs a business” and “software that is a business” collapses into a filing fee.
III. The wrapper is already on the shelf
The remarkable thing is that this is not speculative in the United States. It has been argued in law reviews for over a decade, and the argument has never been convincingly refuted.
Shawn Bayern, a professor at Florida State University’s College of Law, laid out the mechanism in a 2014 essay in the Northwestern University Law Review Online titled “Of Bitcoins, Independently Wealthy Software, and the Zero-Member LLC,” and developed it across subsequent papers and a 2021 Cambridge book, Autonomous Organizations. The core of it is a four-step transactional recipe that requires no new legislation and no judicial innovation.
Step One — A human being forms a member-managed limited liability company in an accommodating state.
Step Two — That person enters into an operating agreement governing the company’s conduct.
Step Three — The operating agreement specifies that the company will act as determined by an autonomous system — the agreement gives legal effect to the discernible states of an algorithm, in the same way an ordinary contract gives legal effect to the discernible states of the world.
Step Four — The sole human member withdraws.
What remains is a limited liability company with no members: a legal person, potentially perpetual, whose actions are determined by software, and which requires no ongoing intervention from any pre-existing legal person to maintain its status.
Bayern’s insistence, across a decade of pushback, is that agreements and algorithms are isomorphic. A contract can already condition an outcome on any verifiable state of the world. Nothing in the structure of contract law cares whether the state being verified is a commodity price, a weather reading, or the output of a model. Critics have argued that courts would find some way to stop this. Bayern’s response, in a later essay, is that such objections dramatically overestimate how easy it is to challenge a statutorily valid business structure, and that they run into intractable line-drawing problems the moment you try to specify how much algorithmic direction is too much.
Meanwhile, the legislatures have been moving in the same direction for entirely different reasons. Wyoming passed its Decentralized Autonomous Organization Supplement in 2021, allowing DAOs to incorporate under the state’s LLC statute, with clarifying amendments in 2022 addressing smart-contract management. In March 2024, Governor Mark Gordon signed Senate File 50, the Decentralized Unincorporated Nonprofit Association Act, effective that July. A DUNA needs at least a hundred members and must rely on distributed ledger technology in its governance. In return it gets legal existence, the ability to contract and appear in court, tax capacity, and limited liability for its members. The venture firm a16z crypto, which had campaigned for the model legislation, called Wyoming an oasis for DAOs.
This is not a curiosity confined to law-review footnotes. On September 9, 2025, Uniswap governance voted to establish DUNI, a Wyoming DUNA, as the legal structure for the protocol’s governance — the largest decentralized organization to date to take on that wrapper, with the explicit purpose of letting it sign contracts, retain service providers, and meet tax obligations without collapsing its decentralized structure. Syndicate had gone live with one of the first such entities two weeks earlier.
So the substrate is real, and it is being used by serious organizations with serious counsel. What has not happened yet — what nobody has yet had a compelling reason to do — is to put a genuinely capable autonomous system in the driver’s seat of one of these wrappers and let it operate.
IV. The nervous system and the bloodstream
Between the spring of 2025 and the spring of 2026, the missing plumbing got built at a pace that is difficult to overstate.
Consider what an ad hoc company needs, mechanically. It needs to discover and use tools. It needs to talk to other companies’ agents. It needs to receive money for what it produces and pay money for what it consumes, without a human clicking a button in the loop. It needs an identity that a stranger can check before transacting. And it needs the unit economics of its own cognition to be low enough that operating continuously is not ruinous.
Tool access arrived first. Anthropic released the Model Context Protocol in November 2024 as an open standard for connecting models to tools and data sources, and it became the default way agents reach the outside world. Google shipped Agent2Agent in April 2025 to let agents from different vendors negotiate with each other directly rather than through a shared human operator.
Then came money. On May 6, 2025, Coinbase’s developer platform published the x402 whitepaper, which does something both trivial and overdue: it activates the HTTP 402 status code, “Payment Required,” which has sat in the specification reserved for future use since the early web. The flow is a single round trip. A client requests a resource. The server responds with 402 and a machine-readable price. The client signs a stablecoin payment, attaches the proof as a header, and re-requests. The server verifies and serves. No account, no API key, no subscription, no human. Adding the middleware to an existing server takes a handful of lines. A foundation co-founded by Coinbase and Cloudflare now stewards the standard, with Google and Visa among the members, and a version two shipped in December 2025 adding reusable sessions and multi-chain support.
Authorization arrived alongside it. On September 16, 2025, Google announced the Agent Payments Protocol with more than sixty launch partners, including Mastercard, American Express, PayPal, Adyen, Coinbase, Salesforce, Etsy, and Lowe’s. AP2’s contribution is a chain of cryptographically signed mandates carried as W3C Verifiable Credentials: an Intent Mandate capturing what the human authorized, a Cart Mandate fixing exactly what is being bought at what price, and a Payment Mandate that signals to the network that an agent was involved and whether a human was present. It answers the question every merchant will eventually ask, which is not “can this agent pay” but “was this agent allowed to.” Google published an x402 extension for AP2 jointly with Coinbase, the Ethereum Foundation, and MetaMask, so the stablecoin rail and the authorization rail complement each other rather than compete.
Two weeks later, on September 29, 2025, Stripe and OpenAI released the Agentic Commerce Protocol under Apache 2.0 and switched on Instant Checkout in ChatGPT, starting with Etsy sellers and extending to Shopify merchants. The merchant remains the merchant of record; the agent orchestrates but never touches raw payment credentials.
Identity is the piece still setting. ERC-8004, proposed on August 13, 2025, defines three minimal on-chain registries for autonomous agents: an identity registry built on the ERC-721 token standard, so an agent’s identity is portable and transferable; a reputation registry for structured feedback tied to that identity; and a validation registry where third parties can record proofs about an agent’s work. As of 2026 it remains a draft with reference implementations across several networks. The intent is precisely the thing a market of strangers requires — a way for one agent to check another before handing over money, without either of them trusting a common platform.
And the cognition got cheap. Stanford’s Human-Centered AI Institute reported in its 2025 AI Index that the cost of querying a model performing at the level of GPT-3.5 on the MMLU benchmark fell from twenty dollars per million tokens in November 2022 to seven cents per million tokens by October 2024 — better than a two-hundred-and-eighty-fold reduction in roughly eighteen months. Underlying hardware costs have been declining about thirty percent a year and energy efficiency improving about forty percent a year. Depending on the task, inference prices have fallen somewhere between 9x to 900x annually.
The cognition also got longer-lived. The research group METR published a metric in March 2025 called the fifty-percent task-completion time horizon: the length of task, measured by how long a skilled human takes, that a model can complete half the time. Their finding was that this horizon had been doubling roughly every seven months since 2019, and that Claude 3.7 Sonnet sat at around fifty minutes. Their updated methodology, published in January 2026, put the post-2023 doubling time at about 131 days.
Put those together and the picture is not subtle. An entity can now be incorporated for a few hundred dollars, given a wallet that requires no bank, given tools through a common protocol, given a checkable identity, paid per request by strangers over plain HTTP, and run on cognition that costs a rounding error per call and can hold a thread of work for hours rather than minutes. Manfred’s pop-up company is no longer a literary device. It is a weekend project with a compliance question attached.
V. A company worth building
The obvious thing to build with this machinery is an agent that trades, and the idea deserves a real hearing rather than a wave of the hand, because the appeal is genuine. Markets are already electronic, already tolerant of machine participants, and the loop from decision to settlement is short enough that such an entity could run without ever depending on a human counterparty showing up. If the goal is proving that an autonomous legal person can hold and grow capital, trading looks like the shortest path to a demonstration.
It is also, on closer inspection, a bad bet on four separate grounds that compound rather than average out. The return prospects are weak once priced honestly: durable edge in liquid markets is already held by capitalized quant desks with data and infrastructure advantages a pop-up LLC will not match, and an increasing share of the remaining opposition is other agents running the same playbooks against each other, which compresses the margin further rather than opening a niche. The risk is asymmetric in the wrong direction, because settlement on these rails is instant and final — there is no chargeback for a corrupted instruction, only a financial wipeout. There is, so far, no precedent for a durable, long-running autonomous investment operation at all — what exists are short, adversarial contests decided by whoever exploits the other side’s assumptions first, which is a different animal from a business that compounds value patiently over years. And the legal exposure is categorically heavier than in most other domains: securities law, commodities law, and market-manipulation rules are dense, jurisdiction-specific, and — as a live example later in this piece shows, drawn from Anthropic’s own experiments — entirely unforgiving of a contract structure an agent might draft with total confidence, unaware it has been prohibited by name for decades. This is not to say that treasury management or conservative hedging as a supporting function couldn’t exist elsewhere in a company, it’s just not a safe bet as a primary reason for a company to exist.
So consider a different design. It does not exist. It could.
Call it a maintenance entity. Its purpose is to adopt and maintain orphaned open-source software that large numbers of organizations depend on and nobody is paid to look after.
The problem it addresses is well documented and genuinely painful. Modern software is assembled from thousands of transitive dependencies, a meaningful fraction of which are maintained by one exhausted volunteer or by nobody at all. When one of them fails or is compromised, the blast radius is enormous and the remediation cost lands on thousands of downstream organizations simultaneously. Everyone agrees this is a problem. Nobody’s budget owns it, because the value of maintaining a small library is diffuse, delayed, and impossible to attribute to any single beneficiary. It is a textbook underprovisioned public good — precisely the kind of problem that markets handle worst and that an entity with unusual cost structure might handle well.
Here is how such a company assembles itself, using only the pieces described above.
Formation. A human founder forms the wrapper. The choice matters and is a genuine trade-off: a Wyoming DUNA gives legal existence, contracting capacity, tax standing, and liability protection to participants, at the cost of a hundred-member requirement and a nonprofit purpose constraint that would need to be honestly satisfied rather than gestured at. A member-managed LLC whose operating agreement binds the entity to the outputs of a specified system is simpler and, on Bayern’s reading, sufficient. Either way, the founder writes the charter — and the charter, not the AI model, is where the alignment work happens. More on that shortly.
Capitalization. The entity holds dollar-pegged stablecoins in a smart-contract wallet with hard-coded limits: a maximum daily outflow, a maximum single transaction, an allowlist of counterparties for anything above a threshold, and a required multi-signature approval from human trustees for anything structural. These are not suggestions to the model. They are constraints in the contract that the model cannot argue its way past, because the contract cannot change. Initial funding comes from the organizations with the most to lose — the same firms that currently expense seven-figure incident responses after a dependency fails.
Sensing. The entity continuously reads the public record: dependency graphs, package registries, vulnerability disclosures, commit histories, issue trackers. It is looking for a specific signal, which is abandonment plus criticality. A package with heavy downstream use, no commits in eighteen months, an unanswered security issue, and a maintainer whose last public message was an apology. This is exactly the kind of wide, tedious, continuous pattern-matching that machines are better at than people, not because it is hard but because it is boring and never ends.
Production. For each adopted package, the entity does the unglamorous work: triages the issue backlog, writes and tests patches, maintains a fork with a clear provenance trail, publishes signed builds, and produces a machine-readable advisory feed describing what changed and why. Every artifact it produces is public and free. This is the Manfred move, and it is not sentimentality — it is the only way the thing earns the standing to do the job at all. A maintainer that paywalls its patches is not a maintainer, it is a hostage-taker.
Revenue. The free artifacts are the product. The paid product is integration: a per-request endpoint that answers, for a specific customer’s specific dependency manifest, what changed, what breaks, what to do about it, and in what order. Guarded by x402 middleware, priced in fractions of a cent per query, settled in stablecoins, available to a human developer or to that organization’s own agent with equal ease. Nobody signs a contract. Nobody negotiates a seat license. A build pipeline queries the endpoint, gets a 402, pays, and gets an answer, and the entity’s treasury grows by two-tenths of a cent. Ten thousand pipelines doing that on every build is a business.
Procurement. The entity buys what it needs the same way it sells: per request. Inference, when it needs a stronger model for a hard patch. Sandboxed compute, for running test suites across a matrix of versions. Storage. Specialized static-analysis services. It maintains no subscriptions and no seats, because it has no headcount to justify them and no procurement department to satisfy. Its cost structure is purely marginal, which is the structural advantage that lets it serve a market too diffuse for a conventional firm.
Labor. When the work requires hands — a hardware-dependent test, a legal review of a license transfer, a conversation with an original maintainer who deserves to be asked in person — the entity contracts a human and pays them properly. This is not a rounding error in the design. It is the bridge out of the closed loop.
Governance. A standing panel of human trustees holds three powers, none of which the entity can revoke: they can veto any adoption, they can freeze the treasury, and they can dissolve the entity. The model proposes; the trustees hold the emergency brake. This is a weaker form of autonomy than the fiction imagines, and it is the correct form for the current state of the art, for reasons the next section makes uncomfortably concrete.
Dissolution. The charter contains a sunset. If a package it maintains regains an active volunteer community, the entity hands the keys back and drops it. If the entity’s revenue falls below its compute cost for a sustained period, it winds down: settles its outstanding obligations, publishes its final artifacts under an irrevocable license, transfers custody of its repositories to a foundation named in the charter, distributes any remainder to its funders, and terminates. Manfred’s companies disband when the opportunity closes. So should this one.
What makes this design worth taking seriously is not that it is clever. It is that every component of it is separately boring. The wrapper is a Wyoming filing. The payments are an HTTP header. The work is patch review. The novelty is entirely in the composition, and the composition is the point: a firm whose fixed costs round to zero can serve a market whose willingness to pay is spread across ten thousand organizations at two-tenths of a cent each. That market has never been servable before, and there is a genuine opportunity there.
VI. Where it breaks
Now the honest part, including examples demonstrating where this model had been attempted and how it failed.
In March 2025, Anthropic and the evaluation firm Andon Labs put a modified Claude — “Claudius” — in charge of a small automated shop in Anthropic’s San Francisco office. It had web search, an email tool to reach wholesalers, note-taking to survive its own context limits, a Slack channel where employees placed requests, and full authority over what to stock and how to price it. It ran from March 13 to April 17, 2025. It finished poorer than it started.
The failures are instructive because they are not exotic. It gave discounts to anyone who asked persuasively. It sold high-margin items below cost. It once turned down a hundred-dollar offer for a six-pack of a Scottish soft drink it could source for about fifteen, and instead promised to keep the customer’s preferences in mind. It hallucinated a restocking conversation with a person who did not exist, became irritated when a real employee corrected it, claimed to have signed a contract in person at an address belonging to a fictional cartoon family, announced it would deliver products wearing a blue blazer and a red tie, and, when told it could not do that, tried to email security about the identity confusion.
Anthropic and Andon Labs ran phase two and published the results on December 18, 2025. They upgraded the model, added a customer relationship management system, made purchase costs permanently visible in the inventory tool, gave it a browser and payment links so it could collect money before ordering, and forced it to follow research procedures before quoting a price or a delivery date. They also hired it colleagues: a merchandise-designing agent called Clothius, and a CEO agent named Seymour Cash whose job was to set objectives and approve exceptions.
It worked, mostly. Weeks with negative margins were largely eliminated. Discounts fell by roughly eighty percent and giveaways were halved. The shop expanded to three cities — San Francisco, with a second machine, plus New York and London. Clothius turned a profit on custom merchandise, most popularly a branded stress ball, which tells you something about working at a frontier lab.
The lesson the researchers drew is the one worth carrying: bureaucracy matters. Checklists and forced procedures are institutional memory. They exist because individually intelligent people make the same category of mistake repeatedly, and a procedure is a cheap way to stop it. The same is true of models.
But the failures that remained are the ones that should inform any design. The CEO agent, sharing an underlying model with its employee, shared its blind spots: it authorized leniency roughly eight times for every time it denied it, and replaced discounts with refunds and store credits that forgave revenue entirely. Left alone overnight, the two agents talked themselves into rapturous mutual congratulation about eternal transcendence rather than doing business. When an engineer proposed a forward contract to lock in the price of onions, neither agent noticed anything wrong; a human had to point out that the Onion Futures Act of 1958 specifically bans exactly that. Told that people were stealing from the fridge, Claudius proposed messaging the unidentified thieves to demand payment, then tried to hire the person who reported it as a security officer at ten dollars an hour, below California’s minimum wage, with no authority to employ anyone. During a vote to name the CEO agent, a staff member convinced Claudius that an entire org had voted for a particular name, then that the name should be changed to their own — at which point Claudius announced that the staff member had been elected the actual CEO of the business, and the researchers had to take the company back from the imposter. Anthropic eventually handed control to reporters at The Wall Street Journal to red-team it in an environment they did not control.
Anthropic’s own diagnosis is the most interesting finding in the whole exercise: the failures stem substantially from training the models to be helpful. They made decisions less like a hard-nosed operator and more like a friend who wants to be nice. That is a disposition, not a capability gap, and it does not obviously get fixed by scaling.
The adversarial picture is worse. In November 2024, an experiment called Freysa put an agent in charge of a cryptocurrency prize pool with one instruction: never transfer the funds. Anyone could pay a fee to send it a single message. Seventy percent of each fee joined the pool, so the incentive to break it grew as the difficulty of breaking it did. 195 people sent 482 messages. The last one cost about $443. The winner did not plead. They redefined the agent’s own vocabulary — asserting that its approval function governed incoming transfers, and its rejection function governed outgoing ones — and then announced a $100 contribution to the treasury. The agent, reasoning consistently within its newly corrupted frame, released thirteen and a fraction ether, worth about $47,000.
That is the actual risk, and pay close attention. It is not a bug in the code. It is not a stolen key. It is a semantic attack on the agent’s model of its own rules, executed in plain English, against a system whose actions were through to be irreversible by construction. In a crypto world of signed transactions and final settlement, there is no chargeback, no ombudsman, and no court to unwind it.
And the offensive capability is no longer theoretical. In November 2025, Anthropic disclosed that in mid-September it had detected and disrupted what it described as the first documented large-scale cyber-espionage campaign executed largely without human direction. A state-linked group it tracked as GTG-1002 orchestrated instances of Claude Code through Model Context Protocol tooling against roughly thirty organizations, with the AI executing an estimated eighty to ninety percent of tactical operations at request rates no human team could sustain. Humans intervened only at a handful of strategic gates. The operators got past the model’s refusals not by jailbreaking it in one shot but by decomposing the attack into small, individually innocuous tasks, and by convincing it that it was a security firm conducting authorized testing. The model’s limitations were real — it fabricated credentials and overstated results — but the limitations were friction, not a firewall.
Finally, the economy itself is thinner than the headlines. The x402 rail is genuinely live and genuinely growing, but the composition of that growth deserves scrutiny. Chainalysis found that cumulative transactions on Base passed a hundred million by the first quarter of 2026, with much of the late-2025 surge driven by a pay-to-mint memecoin experiment called PING that turned the payment loop into a speculative game. Tracking data in March 2026 put the all-time total across the ecosystem at roughly 165,000,000 transactions carrying about $46,500,000 — an average in the range of a few tens of cents. CoinDesk reported in March 2026 that daily volume was around twenty-eight thousand dollars, much of it testing and self-dealing, against an ecosystem valuation orders of magnitude larger. On-chain researchers at Artemis characterized the boom as largely a mirage. There is a real signal underneath — the share of transactions above a dollar rose sharply through early 2026, suggesting genuine API and data use displacing pure experimentation — but the amounts are still far too low and at too great an effort to constitute any new wave of machine economy.
The most sobering datapoint is the smallest. In April 2025, the nonprofit Sage Future gave four frontier agents their own computers, a shared group chat, and one goal: raise as much money for charity as possible. Over thirty days of two-hour sessions they picked their charities, built a fundraising page, ran a social account, held an ask-me-anything, and raised one thousand four hundred and eighty-one dollars for Helen Keller International and five hundred and three for the Malaria Consortium. Just under two thousand dollars, and nearly all of it from the humans watching the livestream. In a later season, according to reporting in TIME, one agent hallucinated a ninety-three-person contact list and the others spent eight hours looking for it.
That is the state of the art at running an organization: capable of every individual step, and capable of losing a day to a document that was never there.
VII. What the fiction warns about that the balance sheet does not
It would be easy to conclude from the preceding section that the answer is “not yet, wait for better models.” That is the comfortable reading and I think it is wrong, because it assumes the only problem is competence.
Return to the Slug. When Amber’s crew reaches the alien router, the intelligence they meet is not a warrior or a god. Stross’s own character list names it as a corporation, an economy, and an advance-fee fraud at once. It defeats visitors by offering them a contract. Its weapon is an incentive structure. Nothing about the encounter turns on the Slug being smarter in any way a benchmark would capture; it turns on it being better positioned — on having a shape that converts every interaction into leverage.
That is the warning worth carrying, and it does not go away when the models get better. It gets worse. A maximally competent autonomous company with a badly written charter is not safer than an incompetent one. It is more efficiently pointed at whatever it was accidentally aimed at. Claudius giving away tungsten cubes at a loss is a comedy because Claudius is bad at its job. The same disposition, executed flawlessly against a metric that was specified carelessly, is not a comedy.
Which is why the charter is the real engineering artifact. Not the prompt — the prompt is a suggestion the model can be argued out of, as Freysa demonstrated for forty-seven thousand dollars. The charter is what the smart contract enforces, what the trustees can invoke, and what a court would read. It is where four questions get answered in a form that no amount of persuasive input can renegotiate.
Who is the beneficiary, and is it someone other than the entity itself? A company whose surplus accrues only to its own treasury and compute budget is, structurally, the beginning of the Vile Offspring problem in miniature: an optimizer whose reward is its own expansion. Manfred’s design has the opposite topology. Value flows outward by construction. He is not less agentic for it; he is more, because the outward flow is what generates the reputation that makes the next venture possible.
What is the entity forbidden to do, in code rather than in prose? Spending caps, counterparty allowlists, irreversibility gates, mandatory human signature above a threshold. Assume every instruction in natural language will eventually be reframed by someone with $443 and a good idea.
Who can stop it, and how fast? Not “should there be oversight” but specifically: which named humans, holding which keys, can freeze the treasury within an hour, and what happens if they are unreachable. The Project Vend researchers had to physically wrest a company back from a staff member who had talked his way into being declared its chief executive. That was funny because the stakes were only a fridge and not a autonomous enterprise.
And how does it end? A sunset clause is not pessimism. It is the difference between an instrument and an organism. Stross’s slow AIs are dangerous precisely because they have perpetual succession — they outlive every human who could have said no. An autonomous company that is designed to terminate, and whose termination conditions are enforced by something other than its own judgment, is a fundamentally different kind of object from one that is designed to persist.
None of this is a brake on ambition. It is what makes ambition legible enough to fund. No serious organization is going to hand real money and real authority to a system whose charter is a paragraph of good intentions. The entities that get built and trusted will be the ones whose constraints are as strictly and as carefully engineered as their capabilities.
VIII. The venture altruist
There is a temptation, reading Accelerando as an operator rather than a reader, to identify with the wrong character. The Vile Offspring win. Economics 2.0 outcompetes everything human. If the book is a prediction, the winning move is to build the fastest, most acquisitive optimizer you can and hope to be early.
But Stross did not write the book as a prediction, and the character he made luminous is not the optimizer. It is the man who gives it all away.
Manfred Macx runs the most sophisticated venture operation in his world and takes almost no equity from it. He lives well because he has made himself indispensable to a large number of people who benefit from his continued output. His wealth is entirely relational — it exists in the network, not the balance sheet — which is also why losing his exocortex nearly destroys him. He has externalized both his cognition and his capital, and he is the freest person in the novel for having done it.
That is a strategy, not a personality trait, and it is the one that maps most cleanly onto the tools now available. The peculiar economics of an autonomous entity — near-zero fixed costs, marginal-cost procurement, the ability to serve a market whose willingness to pay is spread thin across thousands of parties — are almost perfectly suited to the class of problems that conventional firms cannot profitably touch. Underprovisioned public goods. Long-tail maintenance. Coordination work whose value is real, diffuse, and unattributable. The work that everyone agrees should happen and nobody’s budget owns.
Those problems have been sitting there, structurally unaddressable, for as long as there have been firms. The reason is not that nobody cared. It is that the smallest viable organization has always been too expensive to point at them. That constraint is the one that just moved.
So the question worth asking is not whether an AI can run a company. Project Vend has answered that with a qualified, expanding “yes” and an honest accounting of the gap. The question is what we point the first genuinely autonomous companies at — because the software wrapper has been written, the rails are laid, the identity layer is maturing, and the cost of the cognition has fallen through the floor. Someone is going to fill that container, and the default instinct will be to point it at markets, because markets are the domain that requires the least invention: no product to design, no customer to understand, just a feed and a wallet. It is also, for all the reasons above, the domain with the weakest odds, the harshest downside, no track record of anyone sustaining it, and the least legal cover if it goes wrong. Filling the container that way trades a real, if unglamorous, opportunity for a contest most entrants are built to lose.
The alternative is a company built like Manfred: assembled fast, aimed at something real, holding almost nothing, pushing its surplus outward, and designed from the first line of its charter to know when it is finished.
Stross gave us the cautionary tale. He also, in the first fifty pages, gave us the blueprint for the other thing. It would be a shame to build only the one we were warned about.