Accelerators still cast ingots

I. The ingot problem

Accelerators still cast ingots.

For most of the history of steelmaking, you poured molten metal into individual moulds, let it cool, and got ingots: discrete blocks, produced in batches, each one waiting for the next stage of its life to begin. Then you reheated them, paying twice for the same heat, and rolled them into something useful. The industry ran this way for a century, and then it stopped, almost everywhere, within a couple of decades. Continuous casting replaced it: the metal flows without stopping, and each piece is cut to length when it is ready, never when the batch is.

Now look at how venture programmes work. Applications open on a date. Applications close on a date. A cohort is selected, starts together, proceeds together, and finishes together on a stage, in front of investors, on demo day. The industry's own word for it is batch.

Why? The honest answer has nothing to do with pedagogy. A partner can read only so many applications, sit in only so many interviews, hold only so many companies in their head at once. Reading capacity is scarce, so applications get queued; queues get processed in blocks; blocks need calendars. The batch is an artifact of human reading capacity. Everything downstream of it, the application window, the cohort number, demo day itself, is scaffolding around that one constraint.

Which makes it the first thing that stops making sense when the reading is done by agents. An agent-run evaluation accumulates no pile of applications waiting for a partner's attention. It runs when you show up. The queue disappears, and with the queue goes the reason for the batch.

Steel has already made this exact transition once. I decided to make it the model.

II. A silence, declared

I wrote the first version of this essay in August 2026, after fourteen weeks without publishing anything. The silence was more than editorial. The building paused too: the venture repositories had been quiet since spring, and no company had been incorporated. Two ventures were live even so, ai3.studio and Kerubim, both in production, Kerubim with a limited subset of what it will do. A running platform is still a long way from an incorporated company, and the zero stood.

I could have dressed that up. The operating model we had just rebuilt forbids it: one of its written invariants is that a dormant venture is never presented as active, and I see no reason the Foundation should be exempt from a rule it imposes on its ventures. So I said it plainly then, and I leave it here as it was: the plant stopped, and this was the restart.

What broke the silence was a redesign. We took the whole machine apart, how ventures enter, how they get evaluated, how they get funded, how they become companies, and rebuilt it around the argument in section I.

III. The redesign

The old plan had application windows, assessment months, a staffing quarter. All of it is gone, and it was removed on principle: a programme run by agents has no reason to make a human wait for a date.

The process runs continuously, in three movements. The melt is where ideas and people enter and get reduced to what they actually are: evaluation through building, which section V explains. The pour is where capital enters. The cut is where a venture exits as an autonomous company, when it is ready. There is no shared start line and no shared finish line; two founders entering the same day can exit months apart, and the difference in speed is itself information.

The pace is set by the participant. The twelve weeks in our institutional documents were always a ceiling, and got read as a duration; at the speed an agent evaluates, the path from first contact to a decision can be hours. Nobody waits for a window to open, because nothing closes.

We even dropped a word. "Cohort" implies a group that starts together and finishes together, a batch, so we stopped using it, including for ourselves.

One structural note, because it changes what joining means. The programme runs two workshops in sequence. Upstream, private companies get founded and built: real businesses, with founders and cap tables. Downstream, the components those companies prove in production are extracted into the seven Nexi: public protocols, governed as commons. That extraction is the Harvest Model©, which the essay on cathedrals takes at length, and it is the admission criterion: a venture that hands nothing extractable to the commons is a fine business that belongs somewhere else. When you join, you join a company upstream. What your company proves flows downstream, to everyone who builds after you.

IV. The one-person unicorn

Now push the argument in section I one step further, because it does not stop at selection.

The batch is an artifact of human reading capacity. The organisation, the thing with departments, headcount and an org chart, is an artifact of something deeper: human coordination costs. You hire because one person cannot hold sales, legal, finance, operations and engineering in their head at once; and every hire buys capacity at the price of coordination overhead, which is why companies grow meetings faster than they grow output. The org chart is the shape coordination costs force on ambition.

Agents collapse those costs, the way continuous casting collapsed the reheat. So here is the claim, stated as plainly as I can state it: we are heading toward the world of the one-person unicorn, a founder who delegates the business to agents that work for them, across evaluation, legal, growth, treasury and operations, and keeps the two things that do not delegate: judgement, and responsibility. I am not hedging this with "augmented teams" language. The direction is fewer humans per company and more companies per human, and a venture programme that does not plan for it is planning for the past.

Building for that world is the Foundation's actual job, and its contribution has a name. Itabyrium is the infrastructure that makes the delegated company possible: constantly evolving, and offered free of charge, because it exists to serve the mission rather than a revenue line. It is where the harvest lands. The companies founded upstream prove components in production, and what they prove accrues to Itabyrium as commons: what ai3.studio proves about venture creation, what Kerubim proves about verification, what Mercatalia proves about intent matching, becomes infrastructure the next founder gets for free. Every venture through the plant makes the plant better for the one after it.

V. The two reversals

Two things about entering are backwards relative to every programme you know, and they are the two that matter.

You build, and the building is the evaluation. There is no form, no deck, no CV, and no interview panel at the gate. You enter the platform and start working on your idea with the agents; the work itself is assessed as it happens. That is Interaction-as-Assessment©, measured on three scores, with the selection verifiable cryptographically while the content of your work stays private. The barrier to entry is an idea and the willingness to test it. Two consequences are worth spelling out: nobody can charm their way through an assessment that only watches what you do, and nobody leaves empty-handed. If the answer is "not now", you exit with structured work on your own idea, which remains yours. The essay on assessment takes this paradigm apart.

Every company here is born funded. None of the ventures existed as a legal entity when I wrote this, and that stopped being an embarrassment the moment we made it the mechanism, in two steps. Acceptance fixes your terms: role, equity, conditions, set the moment you are in and never renegotiated after. Incorporation waits for one more event: the first letter of intent from an investment fund. No company is created until capital is on its way to it, which means no founder ever holds a stake in an empty shell, and no venture starts life pretending to be funded. You are on the cap table from day one because day one is the day the company begins to exist; in the language of section III, the pour is what casts it. The vehicle is an Italian SRL with innovative-startup status, deliberately structured to convert into an EU Inc when the 28th-regime legislation goes live: Italy now, Europe next, nothing refounded in between.

Which also settles the compensation question honestly. There is no salary, because there is no employer. There is a founding stake in a company that will be born funded or not at all. The capital is a pre-seed of 50,000 to 400,000 USD, and it comes from external investment funds; the Foundation neither invests nor is invested in. What the funds see is your selection dossier, the verifiable scores produced by watching you build, and several funds see it at once, competing on terms. A first-time founder has never had that negotiating position walking into a single office. And I will not hide the other side of the design: if no fund commits, no company is incorporated. That is the price of a system where nothing incorporates on hope, and whether that trade is attractive is exactly the question a founder should be asking. Employees should not apply, and cannot, because the role does not exist here.

VI. One path

The old programme offered five paths into it. This one offers one, and the deletion is the same principle that deleted the batch.

Explorer, developer, contributor: the ladders every ecosystem builds describe ways to work around a company. Read about it, code for it, volunteer for it. They are participation-as-labour, and in the world section IV describes, labour is what gets delegated to agents. Offering you a contributor tier would be offering you the job I just argued out of existence. You can still read everything I publish and push back; this collection is that door, it stays open, and the people who keep showing up with sharp objections get remembered. When repositories go public, you will be able to send a pull request, and a merge will still say more about you than any interview. None of that is a programme, a track or a tier. There is exactly one way to join: found.

What you found has moved since August. The programme is now thirteen ventures in seven groups, A to G, from assessment to treasury, each with a named product, a function stated in one line, and a first functional proof dated in the plan. The ones you have met in this collection sit across the groups: ai3.studio, where an idea is developed with multi-agent AI; Kerubim, AI-powered code writing and formal verification; Mercatalia, connecting supply and demand on multi-species intents; Amarchord, cognitive models for socially assistive agents and robots; Ribalta, multi-species online competitions with verifiable economic mechanisms; K2Yield, automated management of financial resources; and Itabyrium, the infrastructure you met in section IV. A name is a thesis with a commitment attached, and section II already told you how to read the distance between the two.

Each venture opens its founder door only when its preconditions hold: a named product, a defined founder profile, someone available to hold the final conversation, the corporate structure ready. The preconditions are written as checks a stranger could verify, and they go public with every door that opens, so each opening carries its own evidence. There is a bootstrap truth worth stating instead of hiding: the entry system is itself one of the ventures. ai3.studio, in production, is both the front door and a company that still needs a co-founder, and a system that hosts itself has to be lit from outside. So the first founder conversations happen the old way: directly, with me.

VII. Who answers when an agent decides

Every phase of this system has an agentic operator: evaluation, talent, capital, legal, growth. And section IV just argued that whole companies will run this way. So the first objection any serious person raises lands twice as hard: who is accountable when the agent gets it wrong?

Here is the rule, written into the model as an invariant: every agentic decision has a named human who answers for it. For each phase we declare what the agent decides alone and what it must bring to a person. And the final step before any acceptance is a human conversation, with me and with the existing co-founders of the venture you would be joining, because the people already inside have a voice on who builds next to them.

This is no concession to nervous readers. A one-person company delegated to agents is exactly where accountability rots first if nobody designs for it, and a foundation whose flagship venture sells verification of agent behaviour cannot be the place where nobody answers for what agents do. If we get this right, "the first fully AI-native foundation" is a description; if we get it wrong, it is a pose. The invariant is how I intend to keep it a description.

VIII. The boundary, and the first heat

One boundary, declared. Supporting the project financially is real, and it has rules of its own, regulatory and financial, different in kind from the ones that govern founding. It is deliberately outside this essay. What I will say here: everything described in this essay is earned by building, and none of it is for sale.

The founder doors open in the order their preconditions are met, and each one, once open, stays open. That is the practical meaning of replacing the batch: there is nothing to wait for, and there will never again be a deadline to miss.

The metallurgical fact that anchors this whole essay is that a continuous caster is designed to run without stopping: relighting one is the expensive, deliberate act you do intending never to repeat it. Ours stopped before its first pour. We relit it in public, invariants first, and the ventures go through it as the first heat.

The melt is open. Bring something worth casting.