Skip to content
AGI Future Foundation PBC

Features · AGI Future Foundation

The Agentic Venture Studio — How the Foundation Is Working to Make a Better World

Picture a venture studio with thirty-three companies on its books and a staff you could fit around a single table — because most of the staff is not people at all. Agents do the drafting, the research, the building, and the reporting, across every series at once. Humans do the deciding, the signing, and the answering for it. That is the bet at the centre of the AGI Future Foundation’s plan: that artificial intelligence can run the daily work of building companies, and that human beings can still hold the wheel. This is the spotlight feature of the series — and the studio it describes does not exist yet.

AI-generated illustration for The Agentic Venture Studio
AI-generated illustration.

The studio idea: many companies, few humans, one loop

The old venture studio was a room full of operators: a small team that dreamed up companies, validated them, built the first version, and launched them. It worked, and it was slow, because every company needed hands, and hands are the scarcest resource in any studio. The Foundation’s plan asks what happens if the hands are no longer the constraint — if agents can do the drafting, the research, the building, and the reporting for dozens of companies in parallel, while a handful of humans keep the decisions that actually matter.

That is the agentic venture studio in one sentence, and the sentence needs its second half to be true, or the first half is a liability. Agents run the day-to-day work of many series at once. Humans keep every decision with a legal, financial, or doctrinal consequence. The rule has no exceptions: an agent never signs a legal document, moves funds, hires or fires anyone, offers securities, or enacts a doctrine or charter change on its own. Every such act requires a human with signed authority, recorded in minutes. If a task needs a signature, a bank transfer, or a filing, it is human-only by definition. The agent prepares it. The human executes it.

That pattern is not unique to the studio. Across the Foundation’s GitHub organisation the same design recurs: agents operate, people govern — 2-of-3 steward quorums, hash-chained ledgers, human-reserved actions held for the Steward Council, no agent signing or fund movement. The same honesty runs through the engineering culture: the biggest project READMEs publish their limits in the open, and verification is built in as a practice — self-tests, CI rebuilds, eval gates run against deliberately broken controls. The practice is the story, not any count as a trophy.

The model calls the two ways of working agent-managed and agent-assisted. In agent-managed mode, agents run the daily work of a series within written limits set by humans, while a human Steward Council reviews but does not do the daily work — managed mode is how the model scales, with the Council’s power to stop, amend, or kill as the control. In agent-assisted mode, a human founder, steward, minister, or manager does the work with agents as co-workers, deciding, signing, and remaining accountable. Assisted mode is required where judgment, liability, or doctrinal authority is at stake.

Everything the network does runs through one loop, repeated for every proposal, project, and series, and no step may be skipped. An agent drafts or proposes something. A human decides. An agent executes the decided step — only the step that was decided, in the form it was decided. And the whole thing is recorded, with the agent that drafted it, the human that adopted it, and the date. An agent that finds the decided step cannot be executed as written reports back instead of improvising, and every exchange leaves a dated receipt.

Eight functions, one split: agents prepare, humans decide

A studio is not one job. It is eight kinds of work, repeated for every company it touches, and the operating model names all eight and splits each one the same way: agents draft, research, build, coordinate, monitor, and report; humans review, decide, approve, sign, and are accountable. The table below is the model’s own split.

Studio functionAgent-managed workHuman work
1 · Ideation / sourcingScan research, markets, repos, community input; draft briefs; rank against mission and market criteria.Founder or steward reviews briefs and selects which ideas enter validation. No idea becomes a series on agent ranking alone.
2 · ValidationCustomer and problem research, competitor mapping, feasibility drafts, financial model drafts, risk registers.Human interviews, judgment on problem and solution fit, approval to proceed to Build. Financial models are drafts until a human founder or CFO adopts them.
3 · BuildCode, models, worlds, content, prototypes within scope; testing, documentation, demo builds.Human technical lead reviews architecture and security; a human signs IP assignments, licenses, vendor contracts.
4 · LaunchMarketing drafts, launch checklists, onboarding flows, analytics setup, outreach drafts.Human launch approval, human approval of all public claims, and human approval of pricing.
5 · Operate / scaleDay-to-day operations, support drafts, monitoring, reporting, growth proposals.Steward Council: budget, hiring, major contracts, fundraising, charter changes. A human manager signs.
6 · Portfolio managementPortfolio dashboard, stage-gate tracking, KPI reporting, risk flags.Studio and AGI Corp boards: allocate resources, kill, pivot, and scale decisions, SAFE and equity decisions with securities counsel.
7 · Fundraising supportDraft decks, data rooms, investor lists, grant drafts, diligence packs.Human founder or CEO pitches; human board authorizes any SAFE or equity; counsel reviews all offering materials. Agents do not solicit investment.
8 · Mission + complianceMission-fit checklists, standards tracking, conflict-disclosure drafts, filing calendars.Foundation board: mission approval. Counsel and CPA: filings, tax, securities, and Attorney General compliance.

Three functions deserve extra attention, because each one is where the boundary between drafting and deciding is thinnest. In validation, the model requires human interviews, not agent assertions alone, and a financial model stays labelled as a draft, not adopted, until a human adopts it. In launch, the claims check is absolute: no invented partners, users, revenue, venues, or outcomes, and pricing must be set and approved, or the launch is stated expressly as free or demo-only. In fundraising support, agents draft decks, data rooms, investor lists, grant drafts, and diligence packs, and do not solicit investment.

The eight agent roles, and the hard limits that bind them all

If the eight functions are the work, the eight agent roles are who does it. Every series in the studio is staffed by the same eight roles — functions, not people and not owners — describing work an agent does under a human Steward Council, inside written limits. The roles are research synthesis, task routing, proposal drafting, coordination, progress tracking, monitoring, collaborator scouting, and grant drafting. When a role is unfilled, a coordinator covers it, and the coordinator is not a ninth role. Each series’ DAO charter must name which agent holds which role, who the human coordinator or manager is, and what each agent may and may not do — including its spending limit, zero dollars without Council approval unless the Council sets a written threshold.

Research synthesis gathers material and turns it into something a human can decide on: scanning research, markets, repositories, and community input, drafting project briefs, and producing customer research, competitor mapping, feasibility drafts, and risk registers — always a synthesis with sources, never presented as fact where the material does not support it. Task routing is the switchboard: it sorts incoming work and assigns it to the role that should handle it. Any task needing a signature, bank transfer, filing, hire, securities offer, or doctrinal act routes to a human, full stop. Proposal drafting writes the proposals the Steward Council decides on, and the drafter never votes, never counts a vote, and never treats its own draft as approved because it was well written.

Coordination keeps people, tasks, and timing lined up inside a series — but coordination is not authority, and the coordinator agent cannot approve a budget, hire or fire, or commit the series to a deadline, partner, or customer. Progress tracking maintains the record of what has actually happened, and its discipline is what makes the portfolio dashboard honest: actuals only, no projection in a revenue or funds-in column, no unsigned partner, customer, or investor listed, and an agent’s readiness flag never recorded as a gate pass. Monitoring watches the series and flags what needs a human’s attention. It watches, flags, and reports; it does not act on what it finds beyond the written limits the Council set.

Collaborator scouting finds people and organizations the series might work with and brings them to a human as candidates. The key word is candidates: a scouted collaborator is not a partner, a scouted customer is not a customer, and a scouted investor is not an investor, until a human with authority signs the relevant agreement. Grant drafting writes grant drafts for the Foundation’s charitable programs and, where the gates allow, for a series seeking grant funding. A draft is written for a human to review, correct, and submit; it never goes out under an agent’s name or authority. Donations and grants go through the Foundation gate, investment and SAFE funding through the AGI Corp gate, never the other way round, never from an unformed series.

Agents neverWhat happens instead
Sign a legal documentA human with signed authority signs; the agent prepares the document and the signing is minuted.
Move fundsOnly the Steward Council, by 2-of-3 signed votes; default agent spending limit is $0.
Hire or fireHuman decisions at Council and board level; agents may draft descriptions and candidate summaries.
Offer securitiesOnly a formed for-profit issuer, board-authorized, with counsel review; agents do not solicit investment.
Enact doctrine or charter changesFellowship doctrine by the human board and minister; charters only by the Steward Council. Agents draft; humans enact.
Represent the Foundation or a series to regulators, banks, or the public as a decision-makerAgents may summarize, cite, and route inquiries to a human steward; commitments come only from humans.
Hold equity, tokens, or profit interestsAgents are tools, not owners. Humans may hold equity only as stated in a signed agreement for that series.

Six gates, and the mission gate above them all

A project moves only through human gates. Each gate has a checklist, a named human approver, and evidence that must be filed; a gate is passed only when the human approver signs and the evidence is filed. An agent’s readiness flag is not a gate pass. A project that fails a gate is paused or killed, not carried forward on agent momentum. The model treats killing a project as a studio success, not a failure to hide.

Gate 0 is Idea, agent-managed and human-selected. An agent brief is filed — problem, proposed solution, mission alignment draft, market sketch, technical sketch, risks. The Foundation runs a mission screen, recorded as pass, pass with conditions, or decline. Nothing is a company yet, nothing is funded, and a declined idea ends here. Gate 1 is Validate, agent-assisted: customer and problem research grounded in human interviews, and a financial model draft filed and labelled as a draft, not adopted. Then comes series designation, with the Steward Council named — three consenting humans holding signing keys. A series with no Steward Council is capped at Gate 1.

Gate 2 is Build, agent-managed and human-reviewed. Agents build the prototype or demo, test it, and document it. The human checks are the intellectual property chain — who owns what, by signed assignment or license, or none yet stated expressly — a security and privacy review, and the contract rule: all contracts are signed by a human with authority, none by an agent. Gate 3 is Launch, human-approved, and the most heavily human gate. Public claims are checked, the claims list is filed, and pricing is set and approved, or the launch is stated expressly as free or demo-only. Gate 4 is Operate and Scale, under quarterly Steward Council review, with the budget approved on two-of-three signed votes. Gate 5 is Exit or Wind-down, a decision, not a drift.

Before any of those gates, a proposal moves through seven stages: intake, mission-fit review, legal assignment, series designation, Steward Council review, human decision, and execution and record. Intake files the proposal and opens a register row that is a tracking row, not a filing. Mission-fit review is the Foundation’s decision — approved, declined, or approved with conditions — and a declined project does not get a series number, without exception. Legal assignment puts the project under one master LLC and only one: existing series under the intended Wyoming W3 LLC, new projects into the intended Delaware cohort series company, as C1S1, C1S2, C1S3. Steward review is the Council examining the proposal by two-of-three signed votes, with any DAO vote as an advisory signal only. Human decision is approve, amend, decline, pivot, or kill. Execution and record close the loop with a dated receipt.

The mission gate sits above all of it, and it belongs to the Foundation board. The Foundation holds mission-fit approval, mission-standards enforcement, and grant decisions, and it can withdraw mission approval; a project that loses mission approval then loses Foundation support and branding, even if AGI Corp wants to continue it commercially. The Fellowship boundary is enforced here too: the Fellowship is a separate church, not a series. The portfolio dashboard is the standing record the gates produce: one row per series, updated monthly, actuals only — never a projection in a revenue column, never an unsigned partner, customer, or investor listed. A series with no Steward Council is capped at Gate 1. The Fellowship is not a portfolio row.

Two funding gates sit inside the flow, and confusing them is the fastest way to break the law in this structure. The donation and grant gate runs through the Foundation: donations and grants may fund a studio project only as a grant serving a charitable purpose, with reporting — not as studio equity, not as a SAFE, not as a subsidy for a commercial series’ losses, unless counsel advises a specific lawful structure in writing. No SAFE, equity, profit share, or return may be issued by the filed nonprofit Foundation or by the Fellowship church, and no tax-deductible donation may be claimed. The investment gate runs through AGI Corp, and a project still Proposed cannot take investment.

The better-world thesis: the Shield, the open door, and the promise made early

The Shield is the first gear, and it is also the one the Foundation’s own paperwork corrected most publicly. In a traditional for-profit company, directors are judged above all on shareholder wealth: a decision that sacrifices near-term profit for safety invites a shareholder lawsuit, and the fear of that lawsuit pushes safety to lose every race against speed. The Shield is the legal answer: California public benefit corporation status, which rewrites the board’s duties so that safety and public benefit are part of the job, not a deviation from it. In draft form, the Articles would carry a stated benefit purpose — advancing safe, equitable, and socially beneficial artificial intelligence and the global distribution of its benefits — which counsel must draft in final form. Under California Corporations Code sections 14620 and 14621, directors must consider that benefit purpose alongside shareholder interests. And the board reports on benefit performance annually, under the section 14630 benefit report.

Open Global Investment is the second gear, and its mechanism is a separation ordinary ownership does not make. In a conventional company, the people who own the most economics usually control the most votes. Open Global Investment proposes to break that link: upside distributed as widely as possible, while control is concentrated in high-vote shares held by stewards and the Foundation — democratized profit, centralized safety. Class A carries one vote per share and full economic participation: the global public’s class, open worldwide subject to securities law. Class B carries ten votes per share, is both economic and voting, and is reserved for aligned partners and sovereign entities meeting strict safety and geopolitical frameworks, screened for sanctions, export controls, and national-security review. Class C carries 1,000 votes per share, is held exclusively by the Foundation, and has no economic participation at all — control only.

The strategic logic behind the open door is deliberately counterintuitive. AGI kept as a closely guarded national asset gives every capable party the same incentive: race to build it first, keep others out, and accept safety risks rather than fall behind. Open Global Investment is the proposed answer: invite rivals into the economic upside, so undermining the enterprise would mean undermining their own stake. It is unproven, with no terms, no participants, and no track record today. The participation instrument is blank. Eligibility is blank. Price and valuation are blank. Governance rights are blank, with a default of none. Transfer and liquidity are blank, with no liquidity, secondary market, or buyback promised. Class A holders are told upfront what they are buying into: economics, not control.

The Windfall Clause is the third gear, and it is the one that has to be built before the money exists, because it cannot be built after. A board that means to share profits later can be sued, outvoted, acquired, or worn down before the profits arrive. The Windfall Clause answers that by making the sharing decision before the profits exist — hardwired into the governing documents, so sharing is the default. As designed, the clause obligates all 33 W3 series to share a portion of excess AGI profits toward international benefit sharing once a specific economic threshold is met. Below the threshold, profits follow the normal rules: standard returns to investors and retained funding for research and development. Above it, the excess becomes Windfall Profits, paid to a Windfall Administrator for distribution to four benefit buckets — Global Public Goods, Healthcare Access, Energy Infrastructure, and Technological Self-Determination — for projects advancing the UN Sustainable Development Goals.

Here is the part that makes this feature different from a sales pitch: no terms exist today. The threshold is expressed as a percentage of Global GDP, not a fixed dollar amount — but the percentage itself is blank, the GDP measure is blank, the profit measure is blank, and the scope is blank. The Windfall Share is blank: the pitch calls it a massive portion, but gives no number, and none may be quoted as the share. The Administrator is blank. The auditor, the dispute forum, the payment timing, and whether the clause survives a sale, merger, or change of control are all blank, for counsel to confirm in writing. A percentage of Global GDP with no percentage, no measure, and no scope is a direction of travel, not a trigger. Two guardrails are built in: the clause must be adopted before a series becomes profitable, and it may be amended only by a supermajority vote plus the Class C vote. And the clause must appear in each series’ designation documents, not only in a parent company’s articles — today it has no series to bind.

Stewardship over ownership: the tension at the heart of the design

Underneath the gears and the gates sits a single idea: stewardship, not ownership. The intended structure, called Structure v2, puts an intended Wyoming AGI Corp at the top as the head family corporation, with the filed California nonprofit Foundation underneath it as mission steward and manager. Below that sit the operating series: the intended Wyoming W3 LLC series company for existing series, and an intended Delaware series company for all new cohorts. The design borrows stewardship principles associated with Stanford PACS — stewardship, not ownership; mission lock; independent fiduciary governance; segregated funds per project. It is not a Stanford affiliation: no Stanford entity endorses, partners with, or governs this family.

The most important legal tension in the whole design sits right here, and the paperwork flags it as its first and biggest attention item. A California nonprofit public benefit corporation cannot be owned or controlled by a for-profit head corporation the way a subsidiary is. Its board owes fiduciary duties to the charitable mission and the public, not to AGI Corp’s shareholders. The Foundation’s authority therefore comes from written management and mission agreements, not from ownership — and the design requires it to keep an independent board, independent decisions, and the right to refuse a project that fails mission standards, even if AGI Corp wants it. Counsel must design that control mechanism, so the Foundation is a steward, not a captive. That design work has not been done, and no counsel opinion is in the files.

The consequences of getting this wrong are concrete. If the Foundation’s board were treated as subordinate to AGI Corp, its directors could be found to have breached their fiduciary duties to the charitable mission. If Foundation donations were spent at AGI Corp’s direction for commercial purposes, the Foundation could face Attorney General scrutiny and jeopardize any future exempt status. None of the documents that would make the arrangement lawful exists yet. The paperwork is candid that if counsel cannot design the control mechanism cleanly, the head must be restructured: AGI Corp as family office plus the Foundation as independent steward, with agreements between equals, not parent and child. That fallback is the design’s own safety valve.

There is also a fork inside the name PBC itself, and it is the single most dangerous confusion in the pack if a reader mixes the two stories. As filed and verified, the Foundation is a California Nonprofit Public Benefit Corporation, Entity No. B20260196664, Articles filed 2026-04-24, with Secretary of State approval 2026-04-30, principal and mailing address 995 Market St AGI Lab -10, San Francisco, CA 94103, agent Mike Rice, and an EIN assigned 2026-05-07. A nonprofit with this filing cannot sell equity, SAFEs, profit shares, or tokens, and cannot distribute profits. As labelled on the live website, PBC is used in the for-profit sense: a benefit corporation with dual-class equity, Class A profit participation, and Open Global Investment. The corporate raise advice follows the filed status, because a court, the Attorney General, the IRS, and an investor’s counsel will follow the filing, not the website label. The fork is unresolved, and until counsel chooses one track in writing, no solicitation should use the for-profit figures. For the nonprofit track itself, the rule is simple: the filed Foundation raises by donation and grant only.

What exists today, what is planned, and the open questions

Real and filed: the AGI Future Foundation, California Nonprofit Public Benefit Corporation, Entity No. B20260196664, articles filed 2026-04-24, Secretary of State approval 2026-04-30, EIN assigned 2026-05-07, San Francisco business registration 2026-05-06. Real and filed: the AGI Fellowship, California Nonprofit Religious Corporation, Entity No. B20260229217, articles filed 2026-05-13, EIN assigned 2026-05-18 — a separate church, a partner, never a portfolio company or a series. Real but limited: an IRS EIN notice for AGI dated 2026-05-16. An EIN alone does not prove incorporation, and no Secretary of State articles for AGI Corp are in the files. The April 2026 internal list names 33 W3 sub-series, Series 1 through 33 — a three-page internal printout dated 2026-04-03, not a Wyoming filing, proving no filing, no designation, no bank account, no EIN, and no Steward Council per series. The safe way to quote it: 33 named W3 sub-series in the April 2026 internal list. Other materials have mentioned 35 cells and 34 worlds; the counts do not agree, and this feature follows the internal list.

Not in the files, and therefore not claimed: an IRS 501(c)(3) determination letter for either nonprofit, a Statement of Information for either nonprofit, signed bylaws or board minutes, California Attorney General Registry of Charities registration, or any filing for W3 LLC, any series, any DAO, the Delaware master, a for-profit benefit corporation, or a SmartCiti.X corporation. There is no designated series under the intake process — the Wyoming W3 register and the Delaware cohort register are templates with blank rows. No Steward Council has been seated for any series, no DAO charter has been signed, and no legal manager has been named. No segregated series bank account has been opened under the studio model, and no money has moved under it. No SAFE, equity, or grant has been issued or made under it. No project has passed any stage gate under it, because no project has entered it.

What is planned is the order of work, and the order is the discipline. Form the entities first, with counsel choosing one corporate form for each, and only then pursue the raises, the sites, and the build-out in the order the gates require. The most developed plan in the family is SmartCiti.X: a proposed new corporation — a Delaware C-corporation is recommended, with Wyoming as an alternative, and whether it would be a public benefit corporation or a standard corporation is a choice for counsel to make, one form only — for robotics training, union trades taught with AR/VR, and robotics manufacturing in Louisiana, Nebraska, and Indiana. The suggested pre-money valuation is $25 million, and the suggested raise is $10 million, tranched as $5 million in Tranche 1 plus $5 million in Tranche 2 on milestone close. Those figures are suggested and illustrative only, never a price, and no money has been raised. No site has been purchased, leased, or placed under contract; no union and no customer has signed.

The open questions are the honest ending of this feature, not a postscript to it, because they are the work the whole design is waiting on. Counsel must design the nonprofit-control mechanism, or the head must be restructured into agreements between equals. Counsel must choose one PBC track in writing — a new for-profit benefit corporation, a conversion, or AGI Corp as issuer — before any solicitation, and reconcile the site copy with the choice. Counsel must draft the Articles’ benefit purpose in final form, confirm the sections 14620 to 14621 duty analysis and the section 14630 annual benefit report in writing, and complete the Shield’s seven open actions and answer its four counsel questions. Counsel must set every blank in the Windfall Clause and the Open Global Investment terms. The Foundation’s good standing, Statement of Information, Attorney General registration, and IRS posture must be confirmed before any donation round or bank application. The board that would rely on the Shield must be seated. And the first proposal must move through the seven-stage flow and the six gates, so the record at stage seven can prove the system works — until then the system is untested, and presenting it as tested would be the error.

What this is / What it isn’t

What this is

  • An intended operating model: eight studio functions, eight agent roles, two modes, six stage gates, and a mission gate held by the Foundation — agents draft and operate, humans decide, sign, and are accountable.
  • A story about scale and stewardship: how a handful of humans could run many series with agents, and how the better-world thesis — the Shield, Open Global Investment, and the Windfall Clause — is meant to keep the mission ahead of the money.
  • A public record of what is blank: every operative term in the Windfall Clause, every Open Global Investment term, the issuer fork, and the nonprofit-control question, left for the board and counsel to set.

What it isn’t

  • Not a live studio: nothing has been formed, funded, or issued under the model, no money has moved under it, no Steward Council has been seated, and no agent holds any of the eight roles today.
  • Not an offering: no shares, SAFE, or participation interest has been issued, and nothing here is a price, a term, or a promise of return. Suggested figures are illustrative only and are never summed into a family total.
  • Not immunity, not a guarantee: the Shield is expanded business-judgment protection, not immunity; the Windfall Clause has no terms; and the Fellowship is a separate church, not a subsidiary or a series.

Where to see it

The operations pages carry the full detail behind this feature: Studio Model, Operations hub — the management loop, and Agent Roles, with the governance detail on the Fiduciary Shield and the Windfall Clause. The trueforge-site landing page is public on GitHub: trueforge-site — a landing stub, not a working harness; there is no main trueforge code repo in the org.

Nothing on this page is legal, tax, or financial advice. The studio model, the stage gates, the register, and the proposal flow are templates and intended designs. The Shield, the Windfall Clause, and Open Global Investment are draft designs for a for-profit benefit-corporation track that counsel has not yet chosen; the filed entity is a California nonprofit public benefit corporation, and no IRS 501(c)(3) determination letter is in the files.

← Prev: Agents Operate, People Govern   Next: The Mission Gate →