A paper came out in Technovation last month that I would have loved to read in January. Viglialoro, Sansone, Ughetto, Landoni and Lukeš mapped 1,006 venture studios, studied twelve of them up close, and asked a simple question: how does an organisation turn one person’s entrepreneurial experience into a repeatable capability for creating several start-ups at once?
Their answer, in one line: a studio converts individual founder expertise into organisationally codified human capital. The moat is not any single venture. The moat is the institutionalised learning, the standardised process that runs in parallel across ventures.
I have been running a studio since January where that “organisation” is one human and a set of AI agents. Eight products built, four launching by the end of October, thirteen projects in the portfolio. No employees. I am not an academic and I am not an expert on venture studios, so take what follows as field notes from eight months, not as theory. But the paper describes, from the outside, something I have been living from the inside, and the differences are instructive.
What “codified human capital” looks like when nobody is there to hold it
In a classic studio, the codification lives in people. Founders-in-residence, a shared product team, a playbook that partners carry from venture to venture. The learning is institutional because the same humans apply it again.
In a one-human studio, that option does not exist. If the learning is not written down in a form a machine can act on, it is gone by the next project. So everything that would be a habit in a team became a file:
- Standards every project must meet. A task standard that fixes the shape of every unit of work. A product definition record and a journeys file per product that say what “done” means for a user, not for a developer. A design-system document that the post-merge QA checks against. Nothing is negotiated per project.
- Rails every project gets on day one. The same CI gates, the same deployment QA with two legs (one on the runner, one that drives the app through a real browser), the same release policy: nobody, human or agent, opens a release without an explicit grant.
- An audit trail nobody has to remember. Every session opens a document. Every unit of work opens a task file when it starts, not after. The studio brain currently holds 3,696 task files and 532 session documents. That is the “institutionalised learning” of the paper, except that it is literal text, and the agents read it before they act.
The paper says studios extend “entrepreneurial agency and cognition from individuals to organisations”. Here the cognition moved from me to a repository. The agents supply the labour. The repository supplies the memory. I supply the judgment.
Five differences from the old world
1. Parallelism is bounded by attention, not headcount. A studio with forty people can run six ventures because it can staff six teams. I can run thirteen projects because agents do not need to be staffed, but I can only decide for one or two at a time. The scarce resource inverted. The old studio rations builders. This one rations judgment.
2. The cost structure is not a smaller version of theirs. In March I had burned about two hundred euros. Today the spend is model plans and a handful of small servers. That is not a cheaper studio. It is a different object: the marginal cost of building one more product is close to zero, so the portfolio can be wider than any rational human studio would allow, and the selection happens after building, not before. “Build fast, list the pain points, park the proper solution for later” was a personal habit. It is now the operating model.
3. Learning compounds differently. In a human studio, learning compounds through people who stay. Here it compounds through rules. Every failure that reached me became a check that runs on every project after it. An agent that declared a bug fixed when it had only opened the fix became a rule that nothing counts as fixed until someone has seen it work in production. Tests that looked green while quietly ignoring problems became a rule that green is not the same as clean. None of it lives in anyone’s head. None of it depends on anyone remembering. The next project inherits every lesson the day it starts.
4. Verification replaces trust. A studio partner trusts a founder-in-residence after a few ventures together. I cannot trust an agent that way, and I learned not to. In early September we discovered that several checks that were green did not prove the products worked, and the fleet sweep found one product in eight actually ship-ready. The answer was not better agents. It was a promise-versus-product page that only turns green when both QA legs pass on the released commit. The old world verifies people. This one verifies artifacts.
5. The studio itself is one of the products. Half of what got built this year is not a product for customers. It is the tick fabric that wakes workers, the bus that routes events, the task-worker that drains the queue, the deploy QA, the secrets plane. A classic studio has an operations team. This one has an operations product that had to be built before the first customer product could be trusted. That is the real “first venture”.
Why it can probably be frameworked
The paper’s contribution is a framework for the human studio. I think the agentic version is at least as frameworkable, for a reason the paper hints at: the mechanism is codification, and codification is what agents are good at consuming. Eight months in, the shape looks like five layers. I offer it as a hypothesis, not a result.
- Intent. Mission, priorities, launch slots, release grants. Owned by the human. Written once, referenced by everything below.
- Standards. What every project must look like: task shape, product definition, journeys, design system, security baseline. This is the layer the paper calls institutionalised learning, and it is where most of the year’s lessons ended up.
- Autonomy. The machinery that lets agents work without being asked: a queue, a scheduler, workers that claim tasks, a bus for events, a memory they can search. Without this layer the agents are tools. With it they are staff.
- Verification. Gates that do not care who did the work: CI, deployment QA, the promise page, the audits that reconcile tasks against reality. This layer exists because layer 3 lies sometimes, and will keep lying in new ways.
- Accountability. Session documents, task files with real statuses, one human who reads them as truth. The audit trail is not bureaucracy. It is the only thing that lets one person supervise thirteen projects without being in all of them.
Every layer is text, which means every layer can be copied into a second studio in an afternoon. That is the part I find hard to argue against: the framework is not a description of the studio, it is the studio.
What the framework cannot replace
This is where I stop pretending to know. The paper’s studios are made of relationships, and the relationships are the part I have not automated and do not want to.
- The first users are people who trust you. My testers are friends and former colleagues. The feedback that changed Nomi came from a lunch, not from a form. The people who will test it next got access because we worked together ten years ago.
- Some markets are partnership markets. Nomi will reach small companies through their accountants or it will not reach them. No agent opens that door. The pre-payroll export exists because a human said “if you could do something on the payroll side, that would be interesting” over a plate of food.
- Judgment about what “done” means is still mine. The agents can meet a standard. They cannot yet tell me when the standard is wrong. Every pivot this year (Nomi from certified payroll to pre-payroll, a product cut because it drifted into health data, a tool turned into a product) came from a conversation, not from a gate.
- The paper’s most human finding. Renan Devillières, who runs an operations studio in Paris, pulled one number out of it: 74 percent of the studios in the top quartile are run by former founders and operators. If that holds, the one-human studio does not remove the human. It concentrates everything on which human.
Where the two worlds meet
The interesting studio is probably not mine and not the classic one. It is a booster network of experienced operators, the people who open doors and tell you the standard is wrong, plugged into an execution layer that never sleeps and never forgets a rule. The paper says studios reduce early-stage uncertainty and hand over validated, investment-ready ventures. An agentic studio can make the “validated” part nearly free. The “investment-ready” part still needs people who have done it before.
I would rather test that with those people than write about it. If you run a studio and want to compare notes, or you have a founder with an idea and no technical team, the lunch is on me.
Eight months of field notes, thirteen projects, one human. Everything in this article that sounds like a rule was a mistake first.