Most of what my company did yesterday, I didn't do and didn't assign. Software researched the market. Software wrote code, and other software checked it. Something shipped overnight. Technical debt I never logged got paid down because the system that carries it noticed it first. My job was to decide what the company wants and to answer for what it does.
That is the whole job now. And once you've held it for a while, the public conversation about AI starts to sound like people describing a country from its postcards.
AI agents are the loudest topic in software. Watching one hold a job for months at a stretch is still rare: not completing a task — holding a job, with the noticing, the follow-through, and the coordination with other software that a job actually requires. The mainstream still frames all of this as productivity tooling, a faster way for employees to do employee things. That framing misses the actual event. The event is that the employee layer of a company is becoming software, and the company that results behaves differently from anything the last two hundred years of management produced.
I'm going to lay out what that company is, why the numbers point at it, what the next decade looks like if they hold, and what would have to be true for me to be wrong. One thing this is not: a fund thesis. The essays that have defined this subject were all written by investors, and a fund writes a thesis so that someone else will build it. I'm writing this because I already did. Rev9 Labs isn't raising anything and this page asks nothing of you. It's a report.
What an autonomous company is
An autonomous company is a business where software does the work and a person sets the direction. AI agents fill the roles a company would otherwise hire for, organized into teams that plan, execute, verify one another's output, and improve over time. The human decides what the company should do, and answers for what it does. The headcount is software. The direction is human.
The line that matters runs between using AI and being made of it. A company that uses AI hands tools to people, and the tools wait. In an autonomous company nothing waits. And the gap between those two states is not a bigger model or a better prompt. It's architecture. Having built one, I can tell you the load-bearing pieces, because none of them are secrets. They're all sitting in public, scattered across the requests-for-startups pages of the biggest funds in the world, waiting for people to take them literally.
Roles, not tasks. A task automation fires and stops. A role persists. An autonomous engineering team doesn't write code on request; it watches its systems, notices degradation, decides what matters, fixes it, ships. The unit of automation moves from "do this" to "own this outcome," and everything else follows from that move.
Closed loops. Human organizations run open loop. Decisions get made, and results get checked at the next quarterly review, if ever. Y Combinator's Diana Hu describes the alternative in YC's Summer 2026 requests: a business whose activity is queryable by an intelligence layer that compares what's happening against what should be, then adjusts. Observe, decide, act, verify, learn, forever. A company built on closed loops compounds. Improvement stops being a management initiative and becomes a property of the system.
Memory machines can execute against. Every business runs on knowledge scattered through heads, threads, and old tickets. People cope because they half-remember where things live. Agents can't cope, so the autonomous company writes itself down: how decisions get made, how exceptions get handled, what was tried and why it failed. YC's Tom Blomfield calls this the company brain and thinks every company in the world is going to need one. Without it, your agents are brilliant strangers on a permanent first day.
Coordination between teams. One agent is a demo. A company is many teams handing work across boundaries, escalating, negotiating priorities, staying consistent with each other. Andreessen Horowitz's 2026 outlook says enterprises will need systems of coordination layered above their systems of record, managing multi-agent interactions and keeping autonomous workflows reliable. That coordination layer is the org chart of the autonomous company, and it's software too.
A human boundary. Sequoia's Julien Bek splits cognitive work into intelligence and judgment. Intelligence is complex but rule-governed: code to a spec, books to a standard. Judgment is taste and accountability. Models have become very good at the first. The second stays with the person aiming the company. The human does only what only a human can do: want something, choose among futures, and sign for the result.
What the numbers say
The cost of producing software has fallen by ten to a hundred times, by YC partner Jared Friedman's estimate, enough to dissolve the moat of any incumbent whose defense was millions of lines of legacy code. At Cursor, agent users now outnumber users of the tab-completion feature that made the product famous, and the company says a third of the pull requests its own team merges are written by agents working alone in cloud VMs. Jensen Huang says NVIDIA is building toward seventy-five thousand employees working with seven and a half million AI agents — a hundred agents per person, within a decade. Software engineering fell first for reasons that were always structural: the work is digital, the feedback is instant, correctness is checkable. Sequoia's essay this March points at the ratio that frames what comes next: for every dollar businesses spend on software, they spend about six on services, the human work of accountants, brokers, agencies, and outsourced teams. The entire SaaS era competed for the one dollar. The systems I'm describing compete for the six.
None of these numbers has slowed.
The next decade, in three stages
If I can't say where I'd be wrong, I haven't said anything. Here's the sequence as I see it, with dates. And a word about the dates, because they're where readers will want to argue: they mark when each stage becomes ordinary, not when it becomes possible. Every stage below already exists somewhere. The first one runs in my production environment today. A technology arriving and a technology becoming unremarkable are separate events, sometimes decades apart, and most bad forecasting comes from confusing the two.
Now through 2028: the software stage. Autonomous software companies stop being news. Software that maintains itself, fixes its own debt, researches its own market, and pushes its own releases is live; by 2028 it's simply how software gets made, and the median team behind a real product trends toward one person. Dario Amodei put 70 to 80 percent odds — in May 2025 — on the first one-person billion-dollar company appearing in 2026; Sam Altman keeps a betting pool with other CEOs on the year it happens. Whichever year wins the pool, this is the stage where it happens. The quiet story of this stage is plumbing: agent identity, agent-to-agent protocols, machine-readable interfaces, payment rails that let software buy from software. Plumbing eras look dull and build everything.
Roughly 2028 to 2031: the services stage. The autopilots arrive, companies that sell finished work instead of tools. Bek's essay says the next legendary company won't sell accounting software, it will simply close the books, and his playbook starts with outsourced work because swapping a vendor is a purchase order while replacing headcount is a reorg. Bookkeeping, transactional legal, claims processing, compliance, insurance brokerage: intelligence-heavy, already outsourced, verifiable. These go first, and every engagement makes the systems better at the next one, so the climb toward judgment-heavier work is a ratchet, not a wave. Inside ordinary companies, the org chart starts inverting. Fewer doers, more directors, and a new job title, agent supervisor, that a16z lists among the roles of 2026.
The early 2030s: the company stage. The pieces compose. Autonomous teams coordinating with one another becomes a default design for digital-first businesses. Company size decouples from company output for good. The number of companies goes up as the number of people per company goes down, because starting one now costs less than joining one used to. Competition stops being about who hires fastest and becomes about who improves fastest, and improvement belongs to whoever built the better loops.
A second-order business appears somewhere in this stage, and I suspect it shows up early: the company factory. Once autonomous teams are proven parts, assembling them into a new business shrinks from a founding ordeal to an engineering exercise, repeatable the way anything engineered is repeatable. At that point the scarce input isn't capital and it isn't code. It's a person worth amplifying — someone with the judgment to aim a company and the spine to sign for it. The one-person unicorn everyone keeps betting on probably won't be a lone genius who somehow did everything. It will be one person, plugged into machinery built to make one person enough. Which means the strangest labor market of the 2030s will be the market for that person: neither employee nor founder in the old sense, but the director of a company that already knows how to run.
Nothing in this sequence requires new science. The models are sufficient today. The decade belongs to whoever did the work early.
The funds have noticed
Read the 2026 output of the major venture firms side by side and something odd happens: they stop reading like separate predictions and converge on a single company description.
Y Combinator's Summer 2026 requests ask for AI-native service firms ("Instead of giving you a tool, they just do the work"), a company brain for every business, an AI operating system that turns a company into a self-improving loop, software built for agents rather than people, and challengers to every incumbent SaaS product. Five requests. One company. Andreessen Horowitz's Big Ideas 2026 describes agent-native infrastructure for recursive machine workloads, orchestration layers for fleets of digital workers, and the human roles that supervise them. Sequoia has spent two years assembling the case in public, from Konstantine Buhler's agent-economy pillars to Bek's services thesis. Union Square Ventures led a seed round in December 2025 for a startup whose stated ambition is a company run entirely by agents, founder as the only human required. Alibaba.com's president says the one-person unicorn is no longer a theoretical outlier.
Then in May the category got its first spectacle. Polsia, a one-founder, zero-employee company selling an AI system that runs businesses, raised $30 million at a $250 million valuation — and, by the founder's account, let its own agents run much of the raise: the software managed the data room, briefed investors, and handled the back-and-forth on diligence while the founder joined the final calls. The revenue is self-reported, the reviews are mixed, and the skeptics are loud, which is roughly what the first entrant in any new category looks like. Whether Polsia the company endures is beside the point here. The event can't be taken back: institutional investors priced a business with no employees at a quarter of a billion dollars. That is no longer a thesis. That is a term sheet.
It would be easy to read this as a narrative the money is manufacturing. The causality runs the other way. A request for startups is a lagging indicator. Funds write down what the strongest founders are demonstrating in their offices, then ask the market for more of it. When every major firm independently publishes the same request in the same twelve months, they aren't inventing a future. They're writing down a present. I can confirm the present exists.
Where I could be wrong
A thesis that can't fail is a slogan, so here is the honest ledger.
Judgment might resist automation longer than the numbers suggest, and it's woven through more work than the intelligence-versus-judgment split implies. Regulated industries will move at the speed of regulators, not models. Trust is a human artifact; plenty of buyers will pay a premium for a person to blame long after the software outperforms them. The physical world doesn't move at software speed.
And autonomy raises the stakes on error. A person who is wrong is a mistake; a system that is wrong autonomously, at machine speed, is an incident with compound interest. Making software safe to own outcomes is the hardest engineering in the company. Teams that treat it as an afterthought will find out in production. The ones that solve it will look boring from the outside, because production always does.
The nearest falsifier is mine: if Rev9's own loops stop closing, the record of what ships will say so before I do.
The largest uncertainty isn't technical at all. Careers are becoming capabilities, and societies will spend this decade renegotiating what employment and companies are for. I don't have a tidy paragraph that resolves that, and I distrust anyone who does.
If the numbers hold
Notice what the skeptical case has become. It is no longer "this can't work." It is "this will take longer and be messier than the enthusiasts say." That's the argument you make about something that is underway.
If it is: then the defining companies of the 2030s are being architected right now, mostly by small teams and single operators nobody is covering, and they will look strange until suddenly they look obvious. Then the interesting question about any business stops being how many people it employs and becomes how fast its loops close. Then the trillions spent every year on human services work start flowing toward whoever learned, earliest, how to make software accountable.
As for Rev9 Labs: it's the company in the first paragraph. While this was written, its teams were deciding what mattered, handing each other work, checking one another's output, and shipping what passed. The products are public.
Every sentence above will be graded against the date at the top. Good. The headcount is software. The direction is human. The next decade belongs to the companies built that way, and it has already started.
— The operator, Rev9 Labs