Skip to content
AGI Future Foundation PBC

Features · AGI Future Foundation

Agents Operate, People Govern

Nine features into this series, a pattern has been hiding in plain sight. The shield that protects a safety-first board. The watchers that flag threats for a human to judge. The agent economy whose stewards hold the keys. Different projects, different builders, different problems — and the same design underneath all of them: agents do the work, humans hold the decisions. This is the family story, and the rule that holds the family together.

AI-generated illustration for Agents Operate, People Govern
AI-generated illustration.

The pattern that repeats

Start with the pattern, because it is the strongest fact in this whole feature, and it is a fact about design rather than about any single build. Across the Foundation’s repositories and planning documents, the same human-governed-by-design pattern recurs: two-of-three steward quorums, hash-chained ledgers, human-reserved actions, and a hard rule that no agent signs or moves funds. Different teams, different codebases, different purposes — and the same shape underneath. A steward council of three humans, deciding by two-of-three signed votes. Ledgers chained by hashes so that what agents did cannot be quietly rewritten. A list of actions reserved to humans, written down in advance, that no agent may perform no matter how capable it becomes. That recurrence is not a coincidence. It is a design philosophy, applied again and again, and it is the through-line of this entire feature series.

The pattern matters because it answers the question every reader of the previous nine features should have been asking by now: if agents are going to do all this work — watching, transacting, managing, building — then who is actually in charge. The answer the Foundation gives, in repository after repository and document after document, is that humans are in charge, by design, and the design is not a policy statement but a structure. Quorums mean no single human can act alone, and no single agent can act at all on the consequential things. Hash-chained ledgers mean the record of what happened is checkable after the fact, so that a steward’s review is a review of evidence rather than a review of someone’s word. Human-reserved actions mean the boundary is drawn in advance, in writing, rather than negotiated in the moment when an agent asks for more authority. And the no-signing, no-fund-movement rule means the most consequential acts in any organisation — committing it, spending its money — stay with the people who are accountable for them.

It is worth pausing on why that pattern is hard to hold, because the pressure against it is constant and comes from inside the work itself. Agents get more capable, and every increase in capability creates a temptation to widen what they are allowed to do. A system that works well with agents drafting proposals invites the question of why agents cannot approve the small ones. A watcher that flags threats accurately invites the question of why it cannot block the obvious ones. An economy that runs smoothly with agents transacting inside human-set limits invites the question of why the limits cannot be raised, or removed. Each of those questions sounds reasonable in isolation, and each of them, answered yes, moves the boundary a little further from human hands. The pattern survives only if the boundary is treated as a design decision rather than a temporary inconvenience — if the answer to why agents cannot sign is not because they are not capable enough yet, but because signing is a human act, full stop, and capability does not change that.

That is the claim this feature makes, and it is a claim about the family’s intent rather than about its current state. The pattern recurs in designs. Designs are not deployments. The steward councils the pattern requires are, in most cases, not yet seated. The ledgers the pattern requires are, in most cases, not yet keeping records of live operations. The human-reserved actions are written in planning documents, not yet tested by a real agent asking for an exception. A pattern in designs is a promise about how things will be built, and a promise is only as good as the discipline that keeps it when building gets hard. This feature leads with the pattern because it is true about the design, and it will spend the rest of its length showing where the design stands, piece by piece, without pretending the pieces are further along than they are.

The thirty-three: a vision in a list

The largest piece of the family is also the simplest to describe and the easiest to overstate. The W3 network is, in the Foundation’s plan, a Wyoming series company housing 33 named sub-series, Series 1 through 33, each a separate cell with its own assets, books, and liabilities, each DAO-managed day to day with a human Steward Council deciding. The names run from Cognition.X through InfoGurus, covering learning, empathy, diagnostics, biopharma, blockchain, robotics, aerospace, solar, genomics, the metaverse, and more. The count of 33 comes from an April 2026 internal list: a three-page internal printout dated 2026-04-03, not a Wyoming Secretary of State filing. It proves no SOS filing, no legal designation, no bank account, no EIN, no intellectual property, and no Steward Council per series. The safe way to quote it, to a lawyer, a bank, or an investor, is exactly this: 33 named W3 sub-series in the April 2026 internal list.

The list is a vision, and it should be read as one. A list of 33 names is a statement of ambition: the Foundation intends to build across a whole span of human endeavour, from health to energy to learning to cities, with each series a specialised cell in a network that shares governance, mission, and a windfall obligation. The ambition is the point. A network of 33 series is a different kind of enterprise from a single company with a single product: it is an ecosystem, and ecosystems are judged by whether their parts reinforce each other or merely share a letterhead. The governance model is the answer to that judgment: the same mission gate, the same steward pattern, the same windfall pre-commitment running through every series, so that the network is one enterprise in its values even if it is many enterprises in its operations.

But the gap between the list and the reality is the largest in the whole family, and this feature states it plainly. Only 8 of the 33 series have any repository coverage at all. Twenty-five have none. That means for most of the names on the list, there is no code, no repository, no build of any kind in the verified record — only a name in an internal printout and a planned focus on a series page. This feature does not imply code exists where it does not. The series pages on this site are honest about it: every series carries the same status, intended and planned, not formed, funded, or offered. A reader who counts the 33 names and imagines 33 companies is imagining something that does not exist. A reader who counts the 8 with repository coverage and imagines 8 companies is also imagining too much, because repository coverage is not a company either. What exists is a list, a set of pages, a handful of repositories, and a governance design waiting for series to govern.

Two number disagreements in the wider materials are worth naming, because they are the kind of thing that quietly misleads a reader who does not know they exist. The live Foundation site lists a 66-W3 Wyoming Series network and 33 agentic subsidiaries, and other materials have mentioned 35 cells and 34 worlds. Those counts do not agree with each other, and none of them is a filing. This feature follows the internal list — 33 named series — and flags the disagreement rather than picking the largest number. Until the 66 and 35 figures are explained or corrected with a document, they should not be quoted as fact. The same discipline applies inside the AGI Corp worlds collection, where the world counts disagree with each other — 40, 41, and 44 all appear — so this feature quotes no single count there either, and describes the structure instead: six anchored fabrics, and chartered worlds awaiting founder focus.

The six focus-domain DAOs named in the AGI Corp worlds collection deserve the same care. Cancer, Robotics, Ecology, Climate, Education, and Public Health appear there as focus-domain DAOs. They are not W3 series. Only Robotics DAO matches a numbered series — Series 30 — and the other five must not be presented as series. The distinction matters because the series list is a specific, counted, sourced thing: 33 names in an April 2026 internal list. Adding six more names to it from a different collection, under a different label, would corrupt the one count in the family that is properly sourced. The focus-domain DAOs are a separate grouping, in a separate collection, and this feature keeps them separate.

AGI Corp: the head that is not yet a company

At the top of the intended structure sits AGI Corp, and the most important fact about it is a negative one. It has an EIN, assigned 2026-05-16, on an IRS notice, form CP575A, naming a taxpayer called AGI at 111 Maiden Ln Ste 540, San Francisco, CA 94108, with Form 1120 listed. It has no verified incorporation: no articles of incorporation, no entity number, and no Secretary of State filing were in the files reviewed for this site. An EIN alone does not prove incorporation, and this feature does not present it as proof. The name on the notice is AGI, not AGI Corp, and the difference matters: a short name on a tax notice is not the same as a legal corporate name on articles of incorporation, and no document reviewed for this site states the two names as the same entity.

The intended role is large, and it is worth stating in full, because the gap between the role and the verification is the whole story of AGI Corp. In structure v2, AGI Corp is the company at the top: a Wyoming corporation that would head the network, run the venture studio, and manage the companies to the Foundation’s mission statement and standards. The Foundation would manage companies to mission; AGI Corp would be the commercial head that builds and runs them; the W3 LLC series network and the Delaware cohort series would be the portfolio it builds in. The venture studio is an AGI Corp function rather than a separate entity unless counsel advises otherwise, agent-managed and agent-assisted, with humans executing — signing, moving funds, hiring, and offering securities — and agents preparing, across six stage gates from Idea to Exit and Wind-down. The suggested figures in the planning papers — a $20 million suggested pre-money for the AGI Corp studio and a $2 million suggested studio seed — are illustrative only. Nothing has been raised, nothing has been offered, and no SAFE has been issued.

There is a structural tension in the plan that belongs in the same passage as the plan itself. The Foundation, as filed, is a California nonprofit. A nonprofit cannot be owned or controlled by a for-profit head, and its board must remain independent; control would run only through written mission agreements, designed with counsel. A plan that puts a for-profit corporation at the head of a network that includes a filed nonprofit therefore needs a legal design that keeps the nonprofit independent while letting the commercial head do its work — and that design does not yet exist in a verified document. Until counsel designs it, the head in structure v2 is a direction, not a legal fact. The repository copy that describes projects as owned by, powered by, or operated by AGI Corp is not verified legal fact, and this feature treats it as what it is: the repo describes itself that way. A line in a repository is branding and aspiration, not a filing, and AGI Corp’s incorporation is unverified, EIN only.

The worlds collection attached to AGI Corp tells the same story in a different register. The agi-corp-worlds repository is private, and its contents are spatial fabrics, agent swarms, and an RP1 export route. The structure, rather than any single count, is what this feature reports: six anchored fabrics, and chartered worlds awaiting founder focus. The anchored worlds carry verified coordinates — the SmartCiti Holodeck, Cognition, Locator, Biota, Empathic, Diagnostic — and their fabric files are published, though the assets they reference are not all hosted, so they do not render fully yet. The chartered worlds — the 34 named worlds in the collection, including the six PBC focus-domain DAOs discussed above — await founder focus, latitude, longitude, and an endpoint before anything could attach on RP1, where zero worlds are live today. Bringing any world to RP1 remains a proof-of-concept step, authorised world by world, and no world has yet been reported live there. The collection is a map of ambition with a few anchors set and most of the territory still to be claimed.

Fundraising for AGI Corp exists in the Foundation’s papers as drafts, and drafts only: a post-money SAFE draft, a board authorisation, and a term sheet, with all terms blank. No SAFE has been issued, no money has been raised, no investor has been named, and no Form D has been filed. The same discipline applies to the alternative issuer: the for-profit benefit-corporation raise structure, with a suggested $500,000 to $1.5 million first SAFE tranche and a suggested $8 to $15 million cap, is an alternative issuer to AGI Corp, not an additional raise alongside it. Counsel must choose one issuer and one track in writing first, and no choice has been verified. What exists for AGI Corp is an EIN notice, a portal record quoted as reported, planning papers, and a private repository of worlds. What does not exist is articles of incorporation, an entity number, a Secretary of State filing, a board, a bank account, a cap table, a customer, revenue, an issued SAFE, or a filed Form D.

The Fellowship: a partner, not a piece

The AGI Fellowship is the part of the family most often misunderstood, and the misunderstanding is the easiest to correct, because the facts are filed and the site is live. The Fellowship is a separate church: a filed California Nonprofit Religious Corporation, Entity No. B20260229217, Articles filed 2026-05-13, with Secretary of State approval on 2026-05-18, address 111 Maiden Ln #540, San Francisco, CA 94108, agent Mike Rice, and an EIN assigned 2026-05-18. It is partnered with and supported by the Foundation and the other corporations. It is not a subsidiary, not a W3 series, and not part of the commercial structure. A partner is not a subsidiary, a church is not a series, and support is not control, and this feature repeats the distinction because it is the fact most likely to be lost when the family is described as a whole.

The Fellowship’s own site is where its life actually happens, and the live site is the source for its facts in this feature. As deployed on 2026-11-05, the site at www.agifellowship.church carries 40 articles, with doctrine, training, formation, outreach, agents, get-involved pages, faith bridges, videos, and press, and all 39 videos plus 12 images and 3 logos are deployed and live. Joining is steward-gated through [email protected], with no payment taken on the join page and no tax-deductibility claim. That live record is the one this feature uses. The Fellowship repository’s README is stale: it says 15 deep articles and warns that videos will return 404. Both are out of date against the live site, and this feature does not use the README for Fellowship facts. A stale README is a common thing in a fast-moving project, and the honest response is to cite the live site and say so, rather than to repeat numbers that the deployment has already overtaken.

The partnership model, as stated in the Fellowship’s own planning, is support under written agreements, with an independence clause. Its funds are its own, its governance is its own, and its doctrine is its own. Three practical rules follow, stated as hard rules rather than preferences. First, Fellowship funds never go to a SAFE or a for-profit except on counsel’s written advice, and the default is no. Second, agents operate the day-to-day work under human board and minister direction, and they never enact doctrine, sign, or move funds: doctrine is a human, ministerial act, and signing and moving funds are human acts as well. Third, branding and permission are governed by written rules, so that the Fellowship is never presented as a commercial venture and the Foundation is never presented as a church. No IRS 501(c)(3) determination letter is in the files for the Fellowship, so no tax-exempt status and no tax-deductible donations are claimed, and a Statement of Information was due within 90 days of 2026-05-13, with its filing status unverified.

The Fellowship matters to the family story for a reason that has nothing to do with corporate structure. Every other piece of the family is about building, governing, watching, transacting, and managing. The Fellowship is about meaning: what the work is for, who it serves, and what obligations the builders owe to each other and to the world they are changing. A venture studio can decide what to build. A shield can protect the board that decides. A windfall clause can distribute the profits. None of them can say why any of it matters, or what a good outcome would look like in human terms rather than in financial ones. That is the Fellowship’s place in the family: not as a department of the commercial structure, but as a separate church that asks the questions the commercial structure cannot ask of itself, and that keeps the people doing the work connected to a purpose larger than the next milestone. Its separateness is not a flaw in the design. It is the design.

EventAssist: the separate build

EventAssist is the smallest piece of the story and the one most likely to be misfiled, so it gets a section of its own to keep it in the right drawer. EventAssist is a repository under the AGIFutureFoundation organisation, described there as event assistant agents, and submitted on 2026-10-24 as a solo entry to the Patch the Conference Hackathon run by WeMakeDevs. The language around it in hackathon materials — live on stage, built to win — is hackathon pitch language, and this feature does not repeat it as fact. No result or placement is verified for the entry, and no win or prize is claimed here. It was an entry, built by one person, presented at a hackathon, and that is the whole of the verified record for it.

EventAssist is also deliberately separate. It is not a Foundation series and not a Foundation project. It does not sit in the W3 list, it has no series number, and it is not part of the commercial structure described in the rest of this feature. Its separateness is not an oversight; it is the status the record gives it, and folding it into the Foundation’s portfolio would misrepresent both EventAssist and the portfolio. A solo hackathon build about event assistance belongs to its builder and its event, and its connection to the family is that it lives under the same organisation on GitHub and shares the same agent-building impulse — not that it is a venture of the Foundation’s. This feature names it because a family story that omitted it would be incomplete, and places it outside the structure because the structure is not where it lives.

The honest size of EventAssist is also a useful corrective to the scale of the rest of this feature. The 33-series vision, the venture studio, the agent economy, the windfall clause: these are large designs with long horizons. EventAssist is a single build, for a single event, by a single person, in a single weekend’s tradition of hackathon work. Both scales belong in the story. The large designs give the family its direction. The small builds give it its evidence that the direction is buildable at all: that agents can be pointed at a real problem, in a real setting, with real users, and produce something worth demonstrating. A family that only had large designs would be a planning exercise. A family that only had small builds would be a hobby. The combination — large designs, small builds, and an honest account of which is which — is what makes the story worth telling.

The swarm, the studio, and the work ahead

Two ideas that run through the family’s materials deserve a careful paragraph each, because both are easy to inflate and both matter. The first is the Omega swarm. The Foundation’s work is described in places as a swarm: agents operating together, in parallel, coordinating across tasks, rather than a single agent working alone. The Omega-Claw multi-agent framework, named in the Sentinel Stream repository, is the concrete form of that idea in the record: a framework in which multiple agents divide work and coordinate. Beyond that, the details of the swarm as a standing setup are not known in the verified record: not how many agents it has, not their names, not where it runs, not what model or channel it uses, not what it has actually done. This feature does not invent those details. A swarm, in the record, is an idea and a framework, not a documented deployment, and the honest account leaves it there.

The second is the venture studio, which the next feature in this series takes up in full. The studio is an AGI Corp function in the plan: agent-managed and agent-assisted, with eight studio functions and eight agent roles in the planning documents, economics structured but with no terms, and agents holding no equity. Its stage gates run from Idea through Validate, Build, Launch, Operate and Scale, to Exit or Wind-down. Nothing has been formed, funded, or issued under the studio model yet. The studio is where the family’s designs would become companies, and the fact that it has not yet done so is the most important fact about it. A studio that has formed nothing is a plan for formation, and this feature presents it as one, leaving the full story to the feature that follows.

Two things this feature does not present, and says so explicitly, are the AGI Accelerator as an operating residency and an AGI Accelerator Fund as an existing thing. The Accelerator’s Cohort 1 and Cohort 2 dates have passed, and its status, deposit, and guides are unverified; it is not presented here as an operating residency. The Fund appears nowhere else in the verified record, and it is not presented here as existing. Both belong to the category of things the wider materials mention that the verified record does not support, and the rule of this feature series is to leave that category empty rather than to fill it with hopeful prose. A reader who wants the Accelerator’s status will find it when a document supports it, and not before.

What remains, for the family as a whole, is the work list that every feature in this series has been building toward. The issuer and track must be chosen in writing by counsel. The for-profit benefit corporation, if that track is chosen, must be formed. W3 LLC must be filed, and the series designated one by one, with their own bank accounts, EINs, operating agreements, and seated Steward Councils. The Shield’s open actions must be completed. The Windfall Clause’s blanks must be filled. The watchers must be deployed, the agent economy’s testnet record must grow toward production, and the steward bounds must be published and tested. The Fellowship’s Statement of Information must be verified. Each of those is a separate step, with its own document, and none of them is presented here as done, scheduled, or promised by a date, because no date is verified.

The through-line, and what comes next

Look back across the nine features that led here, and the through-line is not a product, a company, or a valuation. It is a rule, applied everywhere the family has designed anything: agents operate, people govern. The Shield protects a board that chooses safety. The Windfall Clause decides in advance what happens to excess profit. The share classes split economics from control. Open Global Investment invites the world to own the upside while the Foundation holds the steering. The watchers flag threats and record commitments for humans to judge. The agent economy gives agents identity and micropayments inside gates held by stewards. The corporate manager makes a company’s management legible so humans can audit it. The series list names 33 ambitions, and the Fellowship keeps the whole enterprise honest about why it exists. Every piece is a different answer to the same question: how do you let agents do the work without letting them become the authority.

The answer the family gives is structural, not rhetorical. It is quorums, ledgers, reserved actions, and a hard line around signing and moving funds. It is the recognition that capability is not authority, and that the moment an agent is allowed to commit, spend, hire, or decide doctrine, the system has stopped being governed and started being run by its tools. The designs in this series are early, thin in places, and honest about their holes. The pattern inside them is the most finished thing the family has, and it is finished in the only sense that matters: it is stated clearly enough to be held to. A future version of this feature will be able to check the pattern against reality — whether the councils were seated, whether the bounds held, whether the ledgers were kept — and that check is the work ahead.

This feature ends looking forward, without predicting dates or outcomes, because prediction is the one thing the record does not support. What can be said is what the work is: form the entities, seat the stewards, publish the bounds, fill the blanks, deploy the watchers, grow the testnet record toward production, and keep the Fellowship separate and the Foundation independent while doing it. The next feature in this series takes up the venture studio, the place where the designs would become companies. The story after that is not written yet, and it will be written by the humans who do the work, with agents preparing and people deciding, one document at a time.

What this is / What it isn’t

What this is

  • The family story: the 33-series vision, AGI Corp, the Fellowship as partner church, and EventAssist, tied together by one through-line — human-governed by design.
  • A pattern worth leading with: two-of-three steward quorums, hash-chained ledgers, human-reserved actions, and no agent signing or fund movement, recurring across the family’s designs.
  • An honest accounting of what is filed, what is intended, and what is separate — with the live Fellowship site, not its stale README, as the source for church facts.

What it isn’t

  • Not a formed family: only the Foundation and the Fellowship have verified filings; AGI Corp is EIN only, the W3 series are intended, and 25 of the 33 series have no repository coverage at all.
  • Not a win: EventAssist was a solo hackathon entry with no verified result or placement, and it is separate — not a Foundation series or project.
  • Not a prediction: no dates, no outcomes, no operating residency, and no Accelerator Fund are claimed here.

Where to see it

The local pages behind this feature are AGI Corp and AGI Fellowship Church, the W3 series hub with all 33 named series, and the Governance hub for the model the family is built inside. The Fellowship’s own live site is www.agifellowship.church — 40 articles, with all 39 videos, 12 images, and 3 logos deployed. EventAssist’s repository is public on GitHub: EventAssist, event assistant agents, a solo entry to the Patch the Conference Hackathon, WeMakeDevs, 2026-10-24.

Nothing on this page is legal, tax, or financial advice. No offering is made here, and no date or outcome is predicted. AGI Corp is EIN only; its incorporation is unverified.

← Prev: The Agent Economy   Next: The Agentic Venture Studio — How the Foundation Is Working to Make a Better World →