Features · AGI Future Foundation
The Biotech Swarm
A single scientist, however brilliant, cannot read a genome the way a reader reads a book. A genome is three billion letters long in a human being, and the letters only make sense in context — which genes are switched on, which proteins they build, how those proteins fold and meet and fail, and how all of that changes in a sick cell versus a healthy one. The work was never going to be done by one mind. It was always going to take a crowd. The Biotech Swarm is the Foundation's bet that the crowd, this time, will be made of software: not one giant artificial intelligence, but a swarm of smaller agents, each with a job, passing work between them, checking each other, arguing in the disciplined way that good science argues, until something emerges that no single agent could have found alone. It is a beautiful idea. It is also, right now, a very early one, and this feature will hold both of those facts in the same hand for the whole of its length.

AGI Future Foundation is a California Nonprofit Public Benefit Corporation, Entity No. B20260196664, Articles filed 2026-04-24, EIN assigned 2026-05-07. No IRS 501(c)(3) determination letter is in the files reviewed, so this feature does not claim tax-exempt status and does not claim that any donation is tax-deductible.
One mind is not enough
Picture a laboratory at midnight, not as a place of glassware and white coats, but as a place of waiting. The machines have finished their runs. The data is in. Somewhere in a server, a file sits that contains the answer to a question somebody cares about — which gene, which protein, which pathway — and the problem is not that the answer is missing, but that it is buried. Buried under volume, under noise, under the sheer number of ways a living system can be read. A human researcher can hold a hypothesis in mind and test it, carefully, one at a time. What a human cannot do is hold ten thousand hypotheses at once, test them all, throw away the nine thousand that fail, and bring back the thousand that survive, ranked, with the evidence attached. That is not a failure of intelligence. It is a failure of bandwidth, and bandwidth is exactly what software has.
The swarm idea is the response to that bandwidth problem, and it is an old idea wearing new clothes. Nature has been running swarms for a very long time: bees that find a flower no single bee could find, ants that build a bridge no single ant could design, immune cells that hunt a pathogen no single cell could catch. The trick, in each case, is the same. No individual knows the whole plan. Each individual follows simple local rules, talks to its neighbours, and the group as a whole does something intelligent that none of its members could do alone. Translate that into software, and you get the picture the genomic-go-platform describes for itself: AI agent swarms for genomics, proteomics, and drug discovery. One agent reads the genome. Another reads the proteins. Another checks the literature. Another proposes a molecule. Another tries to break the proposal. Another keeps the record of what was tried and what failed. The swarm, if it works, is not a single oracle handing down answers. It is a working group, and like any working group it is only as good as its members, its rules, and its willingness to admit when it is wrong.
There is also a human question hiding inside the technical one, and it is the question this feature will keep coming back to: who is the swarm for? A swarm that only serves the people who already have laboratories, datasets, and funding will make the strong stronger, and will leave the rest of the world reading about breakthroughs it cannot afford. A swarm built under a public-benefit purpose, as the Foundation describes its purpose, would have to answer a harder question: how does this help the people whose data it runs on, whose diseases it studies, whose clinics will never see the drug it finds unless the price and the distribution are part of the design from the start? That question has no technical answer. It is a governance question, and it belongs to the human stewards, not to the agents. The swarm can find a candidate. Only humans can decide what a fair way to share it looks like, and that decision has to be made before the candidate exists, not after the value has already been captured by whoever got there first.
Three fields, one problem
Genomics, proteomics, and drug discovery sound like three separate subjects, and in a university they often are, taught in different departments, by different people, with different journals and different conferences. In a living cell they are one continuous story, and the swarm idea only makes sense if you see the continuity. Genomics is the text: the DNA sequence, the inherited letters, the variations that make one person different from another, the mutations that appear in a tumour and not in the healthy tissue around it. It is the most famous of the three fields, and the most easily misunderstood, because a sequence on its own tells you what could happen, not what does happen. A gene can be present and silent, or present and active, or active at the wrong time, or active in the wrong place, and the sequence alone will not tell you which. That is why the second field matters.
Proteomics is the workforce: the proteins that genes build, that do the actual work of the cell, that carry signals, that build structures, that catalyse reactions, that fail when they fold wrong or are made in the wrong amount. Proteins are harder to study than DNA, because they change shape, they interact, they are modified after they are built, and their abundance does not follow neatly from the abundance of the gene that codes for them. A swarm that only reads genomes is reading the recipe book and never tasting the food. A swarm that reads proteins as well is starting to see what the cell is actually doing, right now, in this tissue, under these conditions. The two readings correct each other: the genome says what is possible, the proteome says what is happening, and the gap between them is often where the disease lives.
Drug discovery is the third field, and it is the one where the stakes become concrete, because it is the field that ends in a medicine a person swallows, or an injection a child receives, or a treatment a doctor prescribes. The traditional path from an idea to a drug is long, expensive, and mostly failure: most candidates fail, and they fail late, after years of work. The promise of an agent swarm is to fail faster and cheaper, to test many more candidates in software before any of them reach a bench, to spot the likely failures early, and to bring forward the few that deserve the expensive, careful, human work of real testing. That promise is real, and it is also the promise most likely to be oversold, because software can only test what it can model, and a model of a human body is always a simplification. A molecule that looks perfect in a simulation can fail in a cell, and a molecule that works in a cell can fail in an animal, and a molecule that works in an animal can fail in a person. Each step is a different world, and no swarm, however clever, can skip the steps. It can only make the early steps faster, and make the decision about which candidates deserve the later steps better informed.
The tool on the table: AGI-Biotech-tool
If the swarm is the idea, the AGI-Biotech-tool is the thing you can actually look at, and looking at it honestly is the most useful thing this feature can do. It is a Python repository, and it has 19 open issues. Its own title is biodao.blockchain, which is why this feature cites it under Series 18, BioDAO, but its content is a molecular workspace — proteins, compounds, and drug discovery — so it plausibly belongs to Series 11 Molecular, to Series 18, or to both. This feature notes that overlap and does not merge the two series, which the Foundation's series pages keep separate. One more fact from the review is worth naming, because it is the kind of detail that says more than a slogan: the tool fails its build on any undeclared network endpoint, and it needs no accounts or API keys. That self-verification practice — a build that refuses to reach out to a network address it was not told about, and a tool that asks for no login before it can be used — is the story, and it is a story about discipline, not about scale. Those two facts, stated plainly, tell you almost everything you need to know about where this work stands. Python is the working language of modern biology and modern machine learning, the language in which most of the tools in this field are written, and the choice of Python says that the builders are working inside the field as it actually is, not inventing a new world from scratch. Nineteen open issues says that the work is unfinished, in the open, with known problems still to be solved. An open issue is not a scandal. It is a to-do list that the builders have chosen to publish, and a project with nineteen open issues and no hidden ones is healthier than a project that claims to have none.
What the open issues mean, in practice, is that this is a project at the beginning of its life, not the middle, and certainly not the end. Early-stage software is supposed to have open issues. It is supposed to have rough edges, missing features, and questions that have not been answered yet. The danger is not that the issues exist; the danger is that they are read as proof of failure, or that they are hidden to create an impression of polish, or that they are ignored while the project moves on to announcing new features. The honest reading is the opposite: a project that shows its work, including the work it has not done yet, is a project that can be judged, helped, and improved. A project that hides its work cannot be any of those things. This feature will not list the issues one by one, because their contents were not part of the verified material supplied for it, and listing them from memory or from a guess would be invention. What it will say, and say plainly, is that the count itself — nineteen, open, visible — is the most reliable indicator of maturity available, and it points to early stage.
Early stage also means something specific about what the tool can and cannot be used for today. It cannot be used as a source of verified results, because there are no verified results in the materials reviewed, and no result should be claimed from a tool that is still working through its own open issues. It cannot be used as evidence that the swarm works, because a repository with open issues is evidence that the swarm is being built, not that it has been proven. What it can be used for is exactly what an early-stage open project is for: reading the code, understanding the direction, seeing what the builders think the problem is and how they think it should be solved, and deciding whether to contribute, to test, to criticise, or to wait. That is a real and valuable use, and it is the use this feature recommends. If you are a biologist, a programmer, or a researcher in one of the three fields, the repository is an invitation to look, not a product to buy. If you are not, the repository is a reminder that the most interesting work often starts as a rough, public, unfinished thing, and that judging it by the standards of a finished product misses the point entirely.
Two series, one hard problem: who governs the genome
The swarm does not live in a vacuum. In the Foundation's intended structure, it lives between two planned series, and the relationship between those two series is one of the clearest ideas in the whole W3 plan. Genomic DAO, Series 27, would govern the decisions about genomic data. Genome.Blockchain, Series 28, would keep the record of those decisions. Decisions in one place, records in another: a separation of powers, deliberate and useful, because a body that both decides and records its own decisions is a body that can rewrite its own history without anyone noticing. Both series are named in the April 2026 internal W3 series list, which names 33 sub-series, Series 1 through 33, contiguously, in a three-page internal printout dated 2026-04-03, not a Wyoming Secretary of State filing. The list proves no filing, no legal designation, no bank account, no EIN, and no Steward Council per series, and both series are intended and planned, not formed, funded, or offered. That is the whole of what the list establishes, and this feature will not pretend it establishes more.
Genomic data is unlike most data a series might handle, and the planned focus of Series 27 has to start from that fact, as its series page does. A genome identifies a person and their relatives. It cannot be changed if it leaks. Its interpretation changes as science advances, so consent given for one use may not cover a use invented five years later. A DAO in this lane would therefore be, first and foremost, a consent and stewardship body, not a data marketplace. Its job, as planned, would be to hold the line on three questions: did the person whose data this is agree to this specific use, can that agreement be withdrawn, and does the benefit of the use flow back to the people and communities who contributed? Any design that cannot answer those three questions in plain language would fail the Foundation's public-benefit standard before it started. Those are not abstract questions. They are the questions a person asks before handing over the most personal information they own, and a system that cannot answer them does not deserve the data.
Genome.Blockchain, Series 28, is the counterpart, and its planned role is narrower and more testable: a blockchain record layer that would record, in a verifiable way, the provenance of genomic data — where it came from, who consented to what, who held it, and how it was used. The distinction matters, and it is the first thing to get right. A record layer does not decide; it remembers, and it remembers in a form that can be checked later. But the most important design question is also the most obvious one, and the series page states it plainly rather than assuming it away: genomic data itself must not be put on a public, immutable ledger. A genome is identifying, unchangeable, and shared with relatives, and publishing it, or a fingerprint from which it could be reconstructed, would be a breach that could not be undone. The planned role for a blockchain here is therefore the opposite of a data store: it would record consent receipts, custody events, and access grants — the metadata of stewardship — while the genomic data itself stays in controlled, minimised, revocable storage elsewhere. Any future design that proposes putting genomic sequences, or anything derived from them, on-chain would need to be refused at the mission gate, and this feature records that boundary now so it can be checked later.
Under the Foundation's intended venture-studio model, each W3 series would be DAO-managed day to day, with agents operating within limits set by humans, and a human Steward Council deciding, signing, and moving funds. Agents prepare proposals, analysis, drafts, and coordination; agents never sign legal documents, never move funds, never hire or fire, never offer securities, and never enact doctrine or charters. Humans execute. Every material decision requires a named human's decision and signature. That division is especially important in genomics, because the decisions are the ones that matter most: whether to accept a dataset, whether a consent covers a new use, whether a benefit-sharing term is fair, whether a result should be published or held back. No Steward Council for either series is named in the files reviewed, no operating agreement for either series is in those files, and no agent has authority to sign for, commit, or spend on behalf of either series or the Foundation. The governance described here is the intended model, not a description of an operating body, and anyone reading this feature who wants to help shape that governance should contact the Foundation rather than assuming any of it is already in force.
How would we know if it worked
Every ambitious project needs a way to fail honestly, and the biotech swarm is no exception. In fact, it needs one more than most, because biology is a field where a beautiful idea can be wrong for years before anyone notices, and where the cost of being wrong is measured in wasted laboratory work and misplaced hope. So the question this section asks is simple, and it is the question a sceptical reader should carry through the whole feature: how would we know if the swarm worked, and how would we know if it did not? The answer is not a single test, because a swarm that works across three fields cannot be validated by a single experiment. It is a sequence of tests, each one harder than the last, each one designed to catch a different kind of failure.
The first test is the most basic, and the most often skipped: does the swarm reproduce what is already known, on data it has not been trained on? A system that cannot find the established results, the ones textbooks already agree on, has no business claiming to find new ones. That test sounds easy, and it is not, because it requires the swarm to work on real, messy, independent data, not on the clean dataset its builders used to develop it. A swarm that only works on its own data is a demonstration, not a tool. A swarm that works on other people's data, collected differently, in different places, by different hands, has passed the first gate, and that gate is worth more than any number of impressive demonstrations on home ground.
The second test is about the swarm's ability to say no to itself. Give it a dataset with no real signal in it — a shuffled dataset, a dataset where the labels have been randomised, a dataset where the apparent pattern is an artefact of how the data was collected — and see whether it finds something anyway. A system that always finds something, even when there is nothing to find, is not a discovery engine. It is a pattern-generating machine, and it will fill the literature with false leads that other people will waste years chasing. The swarm needs agents whose specific job is to try to break the result, to argue that the interesting finding is a batch effect, a confounding variable, a statistical fluke, and to demand the extra evidence that would rule those explanations out. A swarm without a sceptic is a chorus, not a science.
The third test is the one that matters most to the people outside the laboratory: does the swarm's output survive contact with the bench? A candidate molecule proposed in software has to be made, tested in cells, and tested again, and most candidates will fail at that stage, however good they looked on a screen. That failure is not a sign that the swarm is useless. It is a sign that the swarm is doing its job at the stage it can do it, and that the later stages are doing theirs. What matters is the rate: does the swarm's shortlist succeed more often than a shortlist made by chance, or by a simpler method, or by a human expert working alone? If it does, and the improvement is real and reproducible, then the swarm has earned its place. If it does not, then the swarm is a faster way of being wrong, and faster wrong is not progress. No such comparison exists in the materials reviewed for this feature, and this feature does not invent one. It states the test so that a future version of this feature can report the result, with numbers, dates, and sources, or report honestly that the test has not yet been run.
The open questions
The second set is about the data, and it is the set that will decide whether the swarm can ever be more than a clever piece of software. Where would the genomic and proteomic data come from, and under what consent, and with what benefit returning to the people it came from? Who decides which datasets the swarm is allowed to use, and who checks that the consent covers the use the swarm wants to make? How is data minimised — holding the least data needed, for the shortest time needed — and how is that minimisation verified rather than merely asserted? What happens when a person withdraws consent: what is deleted, what is retained as a record that consent was once given and then withdrawn, and who verifies the deletion? Genomic data is identifying and unchangeable, and a swarm that treats it as just another input file is a swarm that has not understood its own subject matter. The answers to these questions belong in a written consent architecture, reviewed by counsel and by privacy experts, published for comment, before any real data flows. No such architecture is in the materials reviewed for this feature.
The third set is about the two series and their relationship to the tool. If Genomic DAO and Genome.Blockchain are ever formed, what exactly would each one govern, and how would they relate to the genomic-go-platform and to the AGI-Biotech-tool? Would the tool be owned by a series, or shared across series, or independent of them? Who would decide which questions the swarm works on, and who would decide when a result is ready to be shared, published, or acted on? What would the Steward Councils look like, and who would sit on them, and how would they be chosen, and how would they be replaced? What is the boundary with the neighbouring life-sciences series — BioDAO.blockchain, Series 18, Curable.Blockchain, Series 17, and Diagnostic.X, Series 3 — so that the same data is not governed twice, differently, by two series that never spoke? None of those relationships is a partnership today. Each is a conversation and a written agreement for the future, if the series involved become active, and this feature will not present a future conversation as a present fact.
The last set is the simplest and the hardest: what would success look like, in a way that could be checked? Not a bigger description, not a louder claim, but a checkable outcome. A validation study, with independent researchers, on data the swarm had not been trained on, with the errors reported honestly. A consent architecture, published and reviewed, that answers the withdrawal question before any real data is handled. A shortlist of candidates, produced by the swarm, tested at the bench, with the success rate reported against a fair comparison. Any one of those would be a real step, and a future version of this feature could report it as a fact, with a date and a source. Until then, the work is the questions, and the questions are worth taking seriously, because they are the difference between a swarm that discovers things and a swarm that merely describes itself as discovering them.
What this is / What it isn’t
What this is
- A magazine feature about an ambitious idea: agent swarms working across genomics, proteomics, and drug discovery, with each agent doing one job and the group doing what no single agent could.
- An honest account of an early-stage tool: the AGI-Biotech-tool, a Python repository with 19 open issues, presented as a beginning, not a finished product.
- A description of the genomic-go-platform as the project describes it — AI agent swarms for genomics, proteomics, and drug discovery — attributed to the project, not presented as a verified result.
- Two planned series, Genomic DAO, Series 27, and Genome.Blockchain, Series 28, named in the April 2026 internal W3 series list, intended and planned, not formed, funded, or offered.
- Only 8 of the 33 series have any repo coverage at all, and Series 27 has a landing stub only — so this feature does not imply code exists where it does not.
- A clear governance boundary: genomic data itself must not be put on a public, immutable ledger, and consent must be granular, revocable, and verified.
- A set of validation tests and open questions, stated so a future version of this feature can report whether they have been met.
What it isn’t
- It is not a report of results. No discovery, no candidate, no accuracy figure, no user count, and no deployment is claimed, because none is verified in the materials reviewed.
- It is not a diagnosis, a treatment, or medical advice. No clinical claim is made, and no individual should read this feature as guidance about their own genome or health.
- It is not a finished tool. Nineteen open issues means unfinished work, openly listed, and this feature treats that as the most reliable indicator of maturity available.
- It is not a DAO that exists. No vote, treasury, token, council, or operating agreement for Series 27 or Series 28 is claimed.
- It is not a metaverse world. Neither series has an anchored fabric, and zero worlds are live on RP1.
- It is not an offering, and it is not legal, tax, or financial advice. No investment terms, price, return, or deadline appears in this feature.
Where to see it
The tool: the AGI-Biotech-tool repository, at github.com/AGIFutureFoundation/AGI-Biotech-tool. Start with the open issues — all 19 of them — and with the code itself. That is where the honest state of the work lives, in public, and where a contributor can judge whether to help.
The platform: the genomic-site repository, at github.com/AGIFutureFoundation/genomic-site, where the project describes the genomic-go-platform as AI agent swarms for genomics, proteomics, and drug discovery. The mapping here is ambiguous and this feature keeps it flagged rather than resolving it by guessing: genomic-site is listed for Series 27 Genomic DAO, but Series 28 Genome.Blockchain is also a candidate. Treat that description as the project's account of its ambition, not as an independent verification of its results.
The planned series, on this site: Genomic DAO — W3 Series S27 and Genome.Blockchain — W3 Series S28, which carry the status banners, the filing facts, and the consent and record-keeping boundaries in full. Read the series pages alongside this feature: the feature is the story, the series pages are the record, and the record is deliberately narrower than the story.
Get involved / contact
If this work interests you, the route is a conversation, not a transaction. Write to [email protected] and say who you are, what you do, and what you would like to explore — research, building, review, or a question about the Foundation's mission. A human steward will read it. Please do not send money, do not ask for investment terms, and do not treat this feature as an invitation to invest: there is no offering here, no price, no terms, and no return is promised or implied.
What is genuinely useful at this stage is specific, verifiable help: a named contact in genomics, proteomics, or drug discovery; a written source that corrects or confirms something in this feature; a question that exposes a gap in the plan; or an offer to review code, test the tool on an independent dataset, or help write the consent architecture that this work will need before it touches real data. General expressions of interest are welcome, but they do not create a relationship, a membership, or an obligation on either side. The Foundation's contact address above is the only contact route this feature offers, and no other person is named here as a contact, a founder, a director, or a partner for this work, because no such appointment is verified.
Nothing in this feature is legal, tax, or financial advice, and no signing, membership, payment, investment, or offering is made here. Genomic DAO, Series 27, and Genome.Blockchain, Series 28, are intended/planned, not formed, funded, or offered.