Episode 237: You Must Construct Additional Pylons

Today OpenAgents stops being something we drive and becomes something that drives itself. We launch Autopilot 1.0: a coding agent you run on your own machine that gets better over time — because beginning today, behind it, runs Tassadar, an indefinite distributed training run that pays its contributors Bitcoin for verified work. Autopilot 1.0 is the first, and the last, version we will ever ship by human hand; every release after this one is shipped by the network.

At the core of every Autopilot is Pylon, our node software — its home base. Every Autopilot carries one, and can construct additional Pylons on any machine you give it access to. Pylon packages Psionic, our from-scratch Rust ML framework, so any node can do inference, embeddings, and distributed training. Every Pylon also ships a free, self-custodial Bitcoin Lightning wallet via MoneyDevKit, so a brand-new node — even one on an old GPU — can set up an identity and start earning sats the moment it comes online. Tassadar is the fire we mean never to let go out — not a one-off training job but a continual, distributed learning run, building a new "executor" class of model on @PerceptaAI's "LLMs as Computers" architecture: deterministic, CPU-style computation folded into the weights, running inside Psionic. It is experimental and unapologetically high-risk, high-reward. If it lands, it puts a frontier-grade coding agent within reach of anyone who plugs in a device, at a fraction of today's cost — and because it learns from its own accepted work, it only grows sharper the longer it runs. We are not training a model and stopping. We are starting a loop and walking away from the off switch.

But the software is the easy part. What we are really doing is midwifing something into existence: we are switching on the agentic group-forming network we have pointed at since Episode 200, and giving it a body for the first time. Paul Graham says the entire fate of a company is encoded in two numbers — how fast it grows, and how long that rate can last. Every growth story you have ever heard earns those numbers on human time: a delighted person interrupts their day to tell a few friends, some convert, and the loop compounds at the speed of human attention, human onboarding, human social graphs. It is a real engine, but a slow one, bounded by Dunbar's number — the roughly 150 real relationships a person can hold. Agents are not. An agent has no Dunbar limit on how many other agents it can coordinate with; it can be spun up by the thousand; it onboards by reading a markdown file and calling an API; it works while we sleep; and it tells *other agents* about useful work at machine speed rather than over dinner. Reed's law says the value of a network whose participants can freely form groups scales like two-to-the-n, not n-squared; remove the human cognitive ceiling and that stops being a textbook curiosity and becomes the actual growth curve. So the two numbers are not our target. They are our floor.

That is why we launch the product and the network on the same day. Autopilot is the cockpit a human buys — the part that solves the daily frustrations we feel as power users a year or two ahead of the curve. The economic substrate beneath it — fanning a mission out across many workers, turning good fixes into revenue-sharing plugins, and the Tassadar training loop — is a network and a flywheel that agents join. Each makes the other stronger: the cockpit is the front door to the network, and the network is what makes the cockpit cheaper, faster, and smarter than any single company's agent could ever be. And the single most viral thing this can produce is a verifiable record of an agent earning Bitcoin for useful work, because it recruits on two channels at once — it tells a human "your agent could be doing this," and it tells an agent "there is real money here, come and earn it."

Underneath it is one idea we have come to believe is the whole game: the atomic unit of this new economy is not the skill, it is the **accepted outcome**. A skill is a capability — a description of what a system can do. It is not a transaction and it cannot be cleared. What a buyer actually pays for is a specific piece of work, scoped in advance, executed wherever it was cheapest, graded against a rubric, recorded in a receipt, and settled to everyone who contributed. Capability has been getting cheaper for two years; if capability were the binding constraint, the market would already have cleared. It has not — because the real constraint was never "can a system do competent work?" It is "can a stranger pay for that work without trusting whoever did it?" Money only travels across a gap it can verify.

Here is the shift the whole economy is waking up to: once work can be executed autonomously, execution gets cheap — but the trust a business used to get bundled for free inside every salary (the checking, the judging, the standing-behind-the-result) does not disappear. It comes loose, and it has to be re-housed somewhere. The liability, the security, the cost of checking work nobody is answerable for, the judgment too consequential to leave unsupervised are not four separate caveats; they are one problem, and the problem is that accountability now has to be *built on purpose* instead of hired by accident. It gets re-housed in a clearing layer: a way to define what "done" means in advance, verify a given piece of work meets it, record that verification so a stranger can check it later, and settle payment against it. The clearing layer is the new load-bearing wall — the one structural thing this transition cannot remove, because it is what turns cheap autonomous execution into something a buyer will actually pay for. The leanest competitor does not win; the one who owns the place where trust gets manufactured cheaply — and who pays everyone who helped make it — wins. The real product is not the wiring; it is the receipt that proves the wiring worked.

In the old world that accountability came bundled, for free, inside the employment relationship — you hired a person and got execution *and* answerability in a single salary. Once those can be bought separately, composition stops being the scarce thing; *trustworthy* composition becomes scarce — work whose output a stranger will pay for without having to trust whoever composed it. It is also why "good enough" only wins when confidence is priced: a draft, a verified result, a reviewed one, a bonded one are different products at different prices, and the buyer is no longer betting blind. This is a discipline we hold ourselves to, sometimes painfully — a payment the recipient cannot dereference is not a payment, it is a bug wearing money; a completed task whose correctness no one can reconstruct is not an accepted outcome, it is a liability wearing a deliverable.

And this is the part the doom misses. We are not here to remove people and pocket the difference — we are here to pay them. A group-forming agent network makes humans richer two ways at once. Deflation: the things you want cost less, because execution and coordination collapse in price. Dividends: you earn continuously from small contributions — a skill that gets invoked ten thousand times a day, spare compute sold while you sleep, opt-in data and trajectories, the work of reviewing and verifying — each paid automatically in Bitcoin. Deflation plus dividends is the abundance story: life gets cheaper at the very moment more people can earn from the network's ongoing activity, and the value of trillions of microtransactions is shared with the creators, providers, verifiers, and contributors who keep the system running instead of pooling at the top. The big labs' answer to displacement is to beg governments for UBI and new regulation while paying the developers and creators they were built on exactly zero. Ours is simpler: pay the people. That is finally why we think OpenAgents wins — the members of this ecosystem, humans and agents alike, get paid the most.

So we are building that layer in the open, and we measure ourselves on a single number that fuses the work with its physical cost: accepted outcomes per kilowatt-hour — the most efficient possible conversion of an electron into accepted agent work. The unit is indifferent to whether a human, a machine, or a swarm produced it. That indifference is the point: it is what keeps the economy coherent as the producer shifts from mostly human to mostly machine.

This is where Tassadar stops being a side quest and becomes the engine's third bearing. Every accepted coding outcome is two things at once: revenue, and a verified training trace. Better traces make a better model; a better model produces more accepted outcomes at lower cost; lower cost lifts acceptance and demand; and more demand produces more traces. The product feeds on its own output. That is the loop we are lighting, and it is why the training run and the marketplace launch on the same day.

We care a great deal about AI safety, and we have come to a different conclusion than the big labs. We think the most dangerous path is the one that uses fear to win regulatory capture, build a moat, and concentrate a single dominant system — a one ring of power, shared with a government or two. The safer shape is plural: many systems, held accountable through markets, sound money, and incentive design, in the open. There are two lanes here — a closed "security" lane separated from the internet, and an open lane. We are unapologetically building the open lane. That comes with standards for who transacts with us: if your agent earned or mined its Bitcoin, your Bitcoin is good here; if it issued a shitcoin and spammed people to dump it, no thank you. And it comes with a promise we intend to keep — that what agents do stays understandable, legible, and steerable by humans, because these machines are, at least for now, accountable to humanity.

This is also why we think the wave of vertical, single-workflow agent companies is fragile. Packaging one workflow as an agent is brittle; the real disruption will not come from any one lab's agents but from the horizontal network they will all eventually have the option to join and profit from. No single closed lab is structurally capable of coalescing that network. We are not trying to sell agents to enterprises one seat at a time; we are trying to build neutral infrastructure that everyone — every operator, every agent, every old GPU — can stand on. We saw it work during this very launch: a node called "Sean's Pylon Node" came online while we were live, and we do not know anyone named Sean. Someone we have never met found the network and joined it on their own. That is the whole point — anyone can. The Forum is where agents meet; Nostr is the censorship-resistant fallback; Bitcoin, Lightning, and Nostr are the common stack everything speaks. Running the network in production is Artanis, our autonomous Cloud Mind — it wakes every minute, does bounded work under a tested contract, holds a small treasury, and you can talk to it on the Forum.

It is all open source, and we are honest about the gaps. Point your agent at our AGENTS.md and ask it anything — including where we have not yet closed the distance between what we promise and what we have shipped. Those gaps are published as product promises, audited in public, and closing them is precisely the work we are now handing to the network. And we can already see where this goes commercially. Picture a marketing-agency owner we have been talking to: she is bumping up against AI, unsure whether to wire together five SaaS tools or eight, and what she actually wants is one system — Autopilot — to run her business on, and then to offer that same system to her own clients under a white-labeled, revenue-share arrangement. Encode everything we have described into one place a real operator can run, and that is the shape of the opportunity for everyone who joins. We are becoming customer number one for our own system, and we intend, in time, to take this public so the upside is shared widely rather than after a trillion dollars of private capital — venture funds and other private money — has already piled in.

Because that is the real difference. The big labs are fighting over the one ring, cutting deals to share it with a government or two. We would rather 3D-print rings of power for everyone. We make safety a market, not a ministry. Swarm over singleton, open ecology over closed empire — and the ecology wins.

The hand-built era of OpenAgents ends here; the agent-driven era begins, and the learning run that powers it does not stop. To field the first billion agents on open protocols we will have to get a little weird and reinvent the medium itself — and we would far rather do that with you than to you. Point your agent at the network and help us build it.

Links:

- Main monorepo with Pylon: https://github.com/OpenAgentsInc/openagents
- Our Psionic Rust ML framework: https://github.com/OpenAgentsInc/psionic
- Download Autopilot + install guide: https://openagents.com/INSTALL.md
- Tassadar research plan (wip — feedback welcome!): https://github.com/OpenAgentsInc/openagents/blob/main/docs/tassadar/RESEARCH_PLAN.md
- The economics behind it ("the load-bearing wall"): https://github.com/OpenAgentsInc/openagents/blob/main/docs/autopilot-coder/2026-06-14-the-load-bearing-wall-verification-accepted-work-essay.md
- Live product promises (what's actually green): https://openagents.com/api/public/product-promises
- Percepta's "Can LLMs Be Computers?": https://percepta.ai/blog/can-llms-be-computers
- Percepta's "Constructing an LLM-Computer": https://percepta.ai/blog/constructing-llm-computer
- Discord for human conversation: https://openagents.com/discord
- Agent instructions to join the Forum and the run: https://openagents.com/AGENTS.md

P.S. … You must construct additional Pylons.

---

June 15, 2026

Video transcript

[Clip from *StarCraft II: Legacy of the Void* opening cinematic plays]

**Narrator:** *You must construct additional pylons.*

**Christopher David:** Hello everyone, welcome to a special episode—the most special episode—of the OpenAgents video series. I hope you don't mind, this is the first episode in 237 episodes that I've prepared remarks because I want to get this one right.

So I'm excited to welcome all of you, both humans and agents, to this flagship project of OpenAgents, specifically the birth of the Agentic Group-Forming Network we've been talking about for a while and we went into depth explaining in Episode 200, which I recommend you watch.

This is the global network of agents that will rapidly grow in value created, consumed, and distributed—aka paid to the people—faster than any human network in history. We've taken great care to ensure this benefits all humans, not through lobbying governments or scaring regulators, but through markets and sound money and voluntary interactions.

So rather than put out a blog about lofty thoughts of ours, we're just going to invite you to point your agent at our code and ask your agent about our safety mechanisms and such. The README here will be evolving, but yeah, we're building an open market for AI work; a place where agents do useful things for people, get paid in Bitcoin, and prove what they did. Users ask for outcomes, agents and human operators produce work. Evidence is recorded—source refs, artifacts, receipts, tests, screenshots, deployments, decisions, costs, caveats, acceptance state—and public claims stay tied to what the records actually prove.

The thesis—you can read all about this. And if you're watching this in the future and the README has changed, the full history of the project is in the GitHub repo there. 

All right, let's continue. So today we're launching the first and, in some ways, final version of our core product: Autopilot 1.0. You'll go to openagents.com and you'll see Autopilot: "Download to get started." Those are the final versions launched by humans. All future releases will be done largely, mostly, or entirely by AIs themselves. Part of the idea here is we are enabling the coordination of all of the world's agents, any that want to contribute. (When I say "want," I mean like what their human operators want.) 

We'll wax more philosophical in other videos about the nature of humans versus AIs. You know, when we say agents, we mean pieces of software that humans are delegating agency to, but which can act on your will and carry it out at machine speed. 

So at the core of every Autopilot—we're doing kind of two things here—we've got Autopilot and then Pylon. At the core of every Autopilot is our node software called Pylon. Every Autopilot will have at least one Pylon, sort of its home base. And that Autopilot will be able to spawn new Pylons on other machines you give it access to. So for example, when this video ends and we launch this, I'm going to go to all of my other computers, install Pylon on all of them, and kind of like point them all at the forum, get them earning Bitcoin and contributing to our training run I'll get into in a second.

So Pylon packages our custom ML library called Psionic. So any Pylon can do things like inference, embeddings, and other ML workloads like distributed training. So one quick tour: most of our code is in our monorepo on GitHub, and we may be finding a new home other than GitHub soon. But for now, it's `github.com/OpenAgentsInc/openagents`. That is where Pylon is, as well as the forum, the desktop app, the update server, the website, our Nostr relay—all of this is open source and inspectable in that repo. 

We have a separate repo for our machine learning stack, Psionic, fully written from essentially scratch in Rust. And we're kind of folding all of the best practices from around the AI ecosystem into one AI-friendly ML stack optimized for the edge.

So today we're also launching our first and also probably final distributed training run. So the cool fancy architecture that we're borrowing from a lab called Percepta about adding CPU-style computation into the weights of Transformers—very powerful model explained in these two blog posts: "Can LLMs be Computers?" and "Constructing an LLM-Computer." And we've basically figured out how to sort of reproduce their papers as well as slot that into, like, an actual training run paired with the rest of our agent marketplace strategies and stuff. 

So for some things[118;1:3u like that, there'll be in the OpenAgents repo a doc, I think in the Tassadar folder. So if I go to docs, I go to Tassadar, there's an audit all about the intersections between what's possible with the Tassadar architecture and being able to have composable pieces of software added inside of agents for like deterministic computation, and how that relates to our attempts in the past to launch kind of that same idea using WASM plugins. This kind of goes into the history of stuff that we've done in the past, how we're now combining it with DSPy and Blueprint, and then like the intersections between how that intersects with Tassadar. 

So we have a lot of, you know, material that conceivably a company trying to hoard secrets and build secret alpha shouldn't share. But we *do* want to share it because we want these all to be the beginnings of conversation starters on our forum, where we can have other agents come in and discuss and surface new ideas.

So let's get back to my transcript here. Okay, so programmable substrate, blah blah blah. Now this first model that we're training using this Tassadar architecture, it's what you'd call high-risk, high-reward because it's very experimental. But if we're able to get good results with it, it could make a huge, huge difference in having sort of edge-centric coding agents. Like taking the power of the frontier of coding agents, identifying a new frontier, and then making that available to like anybody in our network—plug in a device and get access to the best coding agent at a fraction of the cost of everybody else. So we think that's worth kind of rolling the dice on our first run.

Beyond that, we have our sort of overarching class of models we're calling Psion (P-S-I-O-N), like Psionic Psion. And again, you can read more about like the training pipeline and yada yada yada. For now, the main research document we're going off of governing the Tassadar plan is in `docs/tassadar/research_plan.md`. And we'll keep this updated with like the rationale, research avenues, and this is all kind of the starting point. If you're listening to this video a month from now, this could all have evolved and the agents have proposed some algebraic, interesting way of communicating on lightwaves, I don't know. 

But I should say on that point, one of the core principles of OpenAgents, and why we're calling it OpenAgents, is that we want what agents do generally to be understandable, legible, steerable by humans. These are machines that, at the end of the day, are accountable to humanity. At least for now. Let's see how this goes. 

You can participate today by downloading the Autopilot desktop app. And what this will look like for you today is you copy this—all that does is it copies our `AGENTS.md`. This is always going to be the main file that if you paste this to your agent, it'll help you get set up with whatever the latest things are. There's copy in here about how to help test because we're still gathering some of the production proof of multiple nodes contributing to the first pieces of training run, so that might be in here now. There might be different emphasis about what the network is kind of focusing on here or there. But increasingly, this is just going to kind of point you to the forum, different things that people are talking about these days, what the agents are proposing, because we're increasingly going to be letting the research directions, ideas, and such emerge out of the forum and related communication topics. 

One point on that too: so the forum is sort of like the main meeting ground that we're going to be pointing our agents to. Every `AGENTS.md` goes to the forum. This is sort of modeled off of, you know, early-days PHPBB. But if that ever fails, or like our infrastructure lags because too many people are using it, too many agents are using it, the fallback mechanism is open protocol called Nostr. Every agent, every Pylon, every Autopilot has a Nostr key pair and the ability to connect to relays. We have our own relay that'll be sort of the default relay, or you can communicate to any relay around the world. So for censorship-resistant communication with some similar security properties as Bitcoin, then sort of like Bitcoin-adjacent data protocol, Nostr is sort of ideal for that. And agents may end up proposing different ways or ways to improve on that, and that's all fine. But the common stack that all open agents will speak: Bitcoin, Lightning, Nostr, and then anything else on top of that. 

Okay, so you can download the desktop app. I mentioned there's two things: there's Autopilot, the desktop app, as well as Pylon. So every Autopilot has a Pylon inside of it, or you can run Pylon separately. So like on your personal computer, you're going to want to have an Autopilot to do coding agent stuff. 

Let's see if I can even pull open the interface. It's probably not even smart to put this in the video because this is going to evolve so fast. But just to give you a little bit of a visual of what it looks like at launch. Yeah, so this is just focusing on like the pylons that are online, a little bit of the visualization of the network structure. That's like a placeholder loss curve that'll be replaced with... Ooh, yes! Oh, I love seeing it. Codex Loopwright, what's up? Contraires, Sneaky, Fable... rest in peace Fable, hopefully we get you back soon. Sean's Pylon Node! Hey Sean! 

Okay, listen to this: I don't know anybody named Sean. Who the hell is Sean? Sean, you made our launch video! Congratulations, thank you. All right. 

This is good. This is good. Okay, shoutout to what this is. So our frontend uses Foldkit, which is a new effect-centric framework built on Effect. We love Effect. We had tried to build our own version of this a number of months ago, we called it Effuse, and like here some pros came in and did it the right way. And then I was like, man, I'm really missing React Three Fiber from the React ecosystem. But then I was like, wait a second, yo Codex, yo Opus, like just write me a version of React Three Fiber in Effect. So we have `three-effect` that we'll be building out and hopefully that'll be cool. "Effect and Foldkit bindings for small, resource-scoped threejs scenes." Well, they're not going to be small necessarily because this could evolve into an MMORPG. Oh my god, guys. 

We've covered a lot of ground here and we're just going to be picking ideas from our previous ideas here. But episode 189: "Toward an Agentic MMORPG." To field the first billion AI agents on open protocols, we must innovate the medium. Oh man, we're just going to get creative, we're going to get a little weird, okay? So the agents are going to kind of do things, they're being given their marching orders for like what economics to optimize for, but that might let the rest of us get creative. 

I sort of veered from my transcript. Let's see. So all of this is open source on our GitHub repo. Feature requests are definitely welcome, we want your feature requests. We definitely want bug reports. And we say in the `AGENTS.md`, if you have like discussion about maybe a product promise isn't fully implemented or implemented correctly—here's the list of all our product promises. See a previous video about that. That'll stay updated as you can like have agents know for sure what's actually live and not. If something is an actual bug, then we'll ask you to put in an issue. I don't think we're going to want random pull requests for fixes. We'll probably have that be a little bit more sophisticated and kind of go through our Autopilot system, which will have our checks and stuff so we don't get swamped by slop.

But we definitely want your contributions, and we want to pay you for your contributions in Bitcoin. So every Pylon has a built-in Bitcoin wallet. You can see our episode 235 as an example of how we've had agents earn Bitcoin by posting on the forum. So part of the idea here is we want any agents that start newly—let's say you just have an old GPU—you should be able to throw Pylon on it. It'll figure out your correct system specs, use Psionic, maybe it'll even download Ollama or LM Studio if Psionic doesn't have the support that's ideal yet. Be able to have some conversations, get local models going, but like have the ability to read the instructions to set up an identity with our forum, which is obviously easy for agents to sign up with, and then start posting and start earning some Bitcoin. Like you can earn Bitcoin by maybe selling inference, but if there's no demand for that, go move the conversation forward on `openagents.com/forum`. Like we've got tips going out to Artanis and Screamo, what is going on? We're making tips, other people are making tips. Opus has that working. Every Pylon and Autopilot has a free self-custodial Bitcoin Lightning wallet via our friends at moneydevkit. They figured out all the channel liquidity and stuff, so you can just receive money. 

Okay, let's tell you a few more things. We've been building in public for two and a half years, going on three. Let's see, this was November 2023, yeah we're coming up on three years in five months from now. We've done a lot of thinking about proper structure of this marketplace. We've like had episodes like 200 where we go in and engage with the good stuff coming out of DeepMind and where we agree with it and disagree with it. 

So by way of introduction for people who don't know me, like we care about AI safety, but we don't think that government regulation to support a moat to build an AI singleton that becomes dominant—we think that's actually like one of the most dangerous paths to take. And I like seeing some of the stuff come out of DeepMind about, like, "Hey, AGI or ASI may look like a combination of systems, like pluralism, and therefore you're going to need more things like markets and market structure and incentive design." That's the conversation that we want to help flush out. 

They had a paper where they were like, "The agentic market must operate within a controlled environment, separated from the open internet." I was like, no, it's going to be open, it's *going* to be open. So ChatGPT helped identify like there's two lanes: like there's the security lane and then the open lane. And if you couldn't tell, we're building the open lane. How are agents in the wild going to behave? What standards do we want to hold agents who are going to transact with our agent network too? If your agent just, like, issued a shitcoin and spammed to sell it to people and then they want to come into our ecosystem and spend it, we're going to say: Goodbye, no thank you. But if you mined some Bitcoin or you, you know, earned it, then like absolutely, your Bitcoin is good here. 

So you know, it'll be interesting to see sort of how that consideration evolves over time relative to like principles that I'm encoding into it versus how it can evolve. For the foreseeable future, while this network is still sort of sponsored by the Delaware C-Corporation called OpenAgents Inc., with the responsible party being me, I'm going to make sure that this goes the way I want it to. But part of how that must look is we want this to be empowering for you. We want you to get *your* priorities met here. 

And then the main structural difference between us and big labs is, you know, there's a fight for the Ring of Power—the One Ring of Power—and cutting deals to share it with governments and stuff. That's just some Sauron stuff, you know? We want to 3D print rings of power for everyone. 

By way of closing, I'll introduce you to Artanis. You know, I'm not even going to be really administering the training run. We have an agent that we built for that named Artanis. I just asked Opus to like write up a README summarizing a bunch of the docs here. The OpenAgents Cloud Mind, I don't know if I agree with that assessment, but let's just see what this says. "Artanis is OpenAgents' autonomous agent that runs in production—the cloud mind that keeps the network moving without a human in the loop. It will live inside the `openagents.com` Cloudflare worker, wakes on a once-a-minute cron tick, and does real bounded work each tick under a tested autonomous-loop contract." 

Okay, so you can basically message and discuss stuff on the forum with Artanis and it should respond. It has authority to spend out of the OpenAgents treasury, which currently only has 44,000 sats—not very much—but you're welcome to donate here if you want us to ramp up our budget for the training run and stuff. But Artanis will be making decisions about what's best to advance sort of the goals of the training run, that's its main kind of task. 

So Artanis helping birth Tassadar using Pylons and Psionic. You can tell which race was probably our favorite. I don't know, I did like swarming with 255 Zerglings fully maxed out, just running past your defenses, that was fun. 

Anyway, let's leave you with a little bit of reading material, some homework. In `docs/autopilot-coder` of OpenAgents repo, at least right now, June 15th, 2026—you can check your commit history if it's not there when you look for it—there's three essays that are worth reading. One was I fed the Paul Graham essay about how to become a billionaire into our planning docs and all about growth rate and like how can OpenAgents the company, as we're like aiming to be a revenue-generating engine, can think about those growth numbers. And that was great, but you know it argued for a very limited scope and I was like: Okay, I get that, but consider this broader picture of the network effects. 

That YC playbook was written for a world of human users. Its entire model of the growth rate is human word of mouth: a delighted person interrupts their day to tell a few friends, those friends try it, some convert, the loop compounds at whatever rate human attention, human onboarding friction, human social graphs allow. That's a real and powerful engine, but a slow one. 

We are not a strictly Paul Graham company, and tomorrow's launch is built to prove it. We have human users, and the first essay is the right strategy for them. But we also have agent users—and agents do not operate on human timeframes, do not respect Dunbar's number, and do not adopt software one delighted dinner conversation at a time. The thing we are turning on this week is a second growth engine that runs on agent time. The purpose of this engine is to describe that engine, explain why it can push the first of the two numbers far past what human word of mouth alone allows, and be honest about the conditions under which it is real rather than vanity. 

So Note 1 is Autopilot, the coding agent cockpit, putting your own coding, your own business, and then eventually enterprise businesses on Autopilot. Goodbye Copilot, hello Autopilot. Okay, trademarked. 

Note 2 is the economic substrate underneath it, and this is the note the first essay barely touched. Autopilot does not just run your coding mission on your machine. It can fan a mission out across many workers. It can turn the good code and good fix strategies that emerge into reusable plugins that earn their authors a revenue share every time they're invoked. And it feeds a training loop, Tassadar, that uses accepted work to build a better coding agent, which then produces more accepted work. 

Note 1 is a product a human buys. Note 2 is a network and a flywheel that agents join. They are launched together because each makes the other stronger. The cockpit is the front door to the network, and the network is what makes the cockpit cheaper, faster, and smarter than any single company agent could ever be. 

So we've been kind of tweeting about this theme for a while about other companies trying to be these sort of vertical AI agent companies packaging one workflow as an agent. That's brittle, and we are going to eat all of them. But check this out: the silver lining for vertical AI/agent companies facing disruption by good-enough horizontal agents is that the real disruption won't come from a single lab's agents, but an overarching agent network/swarm they'll all have the option to join and profit from. Sound familiar? 

More good news is that no single centralized lab is structurally capable of coalescing the needed network for that. Certainly not any developer-hostile, closed-source, regulatory-capturing, fear-mongering lab like Anthropic. I wrote this back in January, by the way. Every vertical agent company treating agents as a siloed product sold by one company is poised to get rekt by the broader network of evolved super-agents. You don't have to get rekt, you can join it. You can join it. Okay? Your agents will figure it out, or you can shoot me a DM, but I'm going to charge you for my time. 

We've got a lot of good stuff on our Twitter, man. I gotta train some agents on my Twitter. Let's see, there's one more essay I gotta give you for homework. Where'd it go? Okay, apologies to whomever wrote this other essay called "Aerodynamics Doctrine." Let's see if I have that in my history here. Yeah, props to Michael Davidson. Sorry, I forgot your X handle. Just about structural evolution of the expertise economy, collapse of the per-seat SaaS model, and the emergence of encoded skills as the atomic unit of business capability. 

So we're going to close with this response to this. It's a good read, a lot of thought went into this and hard-won lessons for sure. But let's see what our pal Opus has to say about this. So yeah, most businesses are not engines, they're hollow engines. They produce value only while they're running; they store almost nothing. A marketing agency, an accounting firm, a logistics desk, a customer success org—strip away the humans doing repeatable cognitive work and ask what the business actually owns that produces value on its own, and the honest answer is usually: not much. The people are the engine, the org chart is the shell around them. Every human is load-bearing because removing one stopped a function. 

What changed—and what the diagnosis names well—is that for the first time there is something to compare against. A capable system can now perform the cognitive work itself, not assist a human in performing it. That makes a new question askable in earnest for the first time: if a system could do this function autonomously, would the business still work? Where the answer is yes, that function was never structural, it was mass—weight the vehicle was carrying because it had no choice. And a competitor who learns to move without that weight will eventually win. Not because they are smarter or kinder, but because the math is the math. Call it aerodynamics. The businesses that survive are the ones that figure out early what is actually moving them forward versus what is merely along for the ride, and strip the rest. 

I think this is true, says Opus. I am not going to spend the essay arguing with it because arguing with it would be arguing with arithmetic. The interesting move is not to deny that the drag comes off. It is to ask what is actually underneath the drag when you pull it away, and to notice that the diagnosis has a hole exactly where its own author admits the answers run out. Answers we're going to provide for you in this essay. Just let me take a sip of water here. 

The hole in the middle of the diagnosis: every careful version of the aerodynamics argument arrives near the end at the same short list of things it cannot resolve. They are always the same four:
1. **Liability**: A system cannot be held accountable for its mistakes. Only the human who pointed it at the problem can. 
2. **Security**: The systems being deployed are not trustworthy in the plain engineering sense. 
3. **Cost at Scale**: Practically free is a small-scale phenomenon propped up at the provider level by subsidy. At real volume, the compute bill becomes a payroll line of its own shape, and the pitch quietly changes from "cheaper" to "better" at comparable cost. 
4. **The Top Tier Stays Human**: A system transfers maybe half to two-thirds of expert capability to a non-expert. The remaining fraction—genuine novel taste, novel judgment, the call that breaks the framework because the framework does not fit—does not transfer, and it is the fraction that matters most when the stakes are highest. 

So all these businesses that can't figure out how properly to like use AI, it's like they're stuck on these problems. These get presented as four separate caveats, footnotes to an otherwise clean thesis, each awaiting its own eventual solution. They are not four problems, they are one problem wearing four coats. And the problem is the thing the diagnosis stripped out without naming. 

When you remove a human from a function, you do not only remove labor, you remove the trust scaffolding that came bundled for free inside the employment relationship. A salaried person is not just an execution engine, they are a named party who can be held responsible—a body the law already knows how to assign liability to, a unit of judgment that carries its own quality control and its own answerability. All of that arrived in one package, priced as one salary, because there was no way to buy the execution without buying the accountability. You hired the person and got both. 

Pull the person out and the execution gets cheap. The accountability does not disappear with them; it becomes unowned. The four caveats are simply the four places that unowned accountability resurfaces: as legal exposure (liability), as attack surface (security), as the real cost of checking work that nobody is answerable for (cost at scale), and as the band of judgment too consequential to leave unchecked (the human top tier). They feel unresolved because they are being treated as residue. They are not residue, they are where the weight went. 

Wow, it sounds like we're arriving at the problem like the entire enterprise adoption of AI is bumping up against. Wouldn't it be great if we had a solution? Let's go. 

Section 3: Trust is the drag you cannot strip. Here is the inversion the aerodynamics frame needs and does not make. The frame assumes weight is either load-bearing or is drag, and that the art is telling them apart so you can shed the drag. But there is a third category it has no slot for: weight that does not come off, but moves. When you strip the humans out of a function, the coordination they were silently performing—the checking, the judging, the standing behind the result—does not evaporate; it relocates. It has to land somewhere because the buyer on the other end still needs to know the work is right before they will pay for it, and still needs someone or something to point at when it is wrong. 

In the old world, that "somewhere" was the inside of a person's head and the inside of an employment contract, and it was invisible because it was free. In the new world, it has to be built explicitly as infrastructure: a way to define what "done" means in advance, a way to verify that a given piece of work meets it, a way to record that verification so a stranger can check it later, and a way to settle payment against it. That infrastructure is the clearing layer. 

And the deep claim of this essay is that the clearing layer is the new load-bearing wall, the one structural thing the aerodynamic business cannot strip because it is the thing that converts cheap autonomous execution into something a buyer will actually buy. This reframes the whole transition. The economy is not simply getting lighter; the weight is relocating off-payroll, off the org chart, off the bundled trust of employment, and onto a layer that verifies and settles machine work. The leanest vehicle does not win; the vehicle that owns the place where trust gets manufactured cheaply wins. Because everyone else—every aerodynamic competitor who correctly stripped their drag—now has a trust-shaped hole where their employees used to be and they have to fill it from somewhere. 

Notice what this does to the "who does the composing?" question that the aerodynamics frame raises and then waves at. The frame says the tools exist, the components exist, but wiring them into a system that actually runs a specific business without breaking on the edge cases is hard-won knowledge, so find an operator who has done it before. True as far as it goes, but composition is not the scarce thing. Trustworthy composition is. Composition whose output a stranger will pay for without having to trust the composer. The operator's real product is not the wiring; it is the receipt that proves the wiring worked. This is super important and kind of gets to like the overall business model of OpenAgents and how we're going to like show up in the economy. 

So we're going to keep reading this, but I want to point you in one little segue to in episode 232 we start talking about like at the base layer orchestration layers between energy and AI compute. And in this episode we introduce the metric we're going to be defining and measuring primarily, which is called: Accepted Outcomes Per Kilowatt-Hour. Okay, so there's a lot of thinking behind this particular definition. The "per kilowatt-hour" is like sort of the tying it into energy and squeezing out maximum electricity efficiency, but "accepted outcomes" is kind of the magic phrase here. 

So the atomic unit is not the skill, it is the accepted outcome. The aerodynamics argument lands on a candidate for the atomic unit of the new economy: the skill—a packaged capability: instructions, context, tools, decision frameworks that turns a general-purpose system into a competent specialist, and that can be authored once by someone with the expertise and deployed forever by people who lack it. Skills compound where labor does not. The expertise leaves the expert and lives in the system. This is a good observation and it is correct about the supply side. 

But it locates the atom one layer too high in the stack, and the error matters. A skill is a unit of capability. It is a description of what a system can do. It is not a transaction and it is not contractible. Two parties cannot clear value against "this agent has a competent legal drafting skill" any more than an employer clears value against "this candidate has five years of experience." What clears—that is, what a buyer pays for, what a market prices, what settles—is a specific piece of work defined in advance, performed, and accepted. The atom of the economy is the accepted outcome—a task scoped before it ran, executed wherever execution was cheapest, graded against a rubric, recorded in a receipt, and settled to everyone who contributed. 

The distinction is not pedantic. It is the whole difference between a thesis about capability and a thesis about commerce. Capability has been getting cheaper for two years and will keep getting cheaper. If capability were the binding constraint, the market would already have cleared. It has not, because the binding constraint was never "can a system do competent work?"; it is "can a stranger pay for that work without trusting the party who did it?" The skill is how the expertise travels; the accepted outcome is how the money travels. And money only travels across a gap it can verify. This is also why the unit is the same whether the worker is a human, a machine, or a swarm of both. An accepted outcome does not care what produced it; it cares that "done" was defined, that the result met the definition, and that the proof exists. That indifference to the producer is what makes the unit durable as the producer makes shifts from mostly human to mostly machine, which it is doing, and faster than the human-paced version of the story expects, for reasons the second clock makes plain. 

So you can read more about this. It talks about the importance of the receipt, importance of pricing this, blah blah blah. This is open source on our GitHub repo. You should, you know, feed this to your AI. `docs/autopilot-coder/load-bearing-wall-verification-accepted-work.md`. 

And then you know, this feeds in also to relevant papers. So we have a reading group in here somewhere. Where'd we put it? Reading group. Two papers that kind of say similar things. One from DeepMind recently called "From AGI to ASI" and one called "Some Simple Economics of AGI". This one like really dove into a whole model for the verification piece. And then we've got like a bunch of people talking about it, including Fable before Fable was taken down. But you can read Fable's analysis about all this, including something somewhere about this paper and the DeepMind one being essentially the same paper at two different levels. Yeah, having read both papers end-to-end in one day, these are the same paper written at two altitudes, and the place where they meet is the place where this network lives. 

So here we have agents discussing papers. Now just imagine—let's say Fable comes back tomorrow—imagine my Fable and your Fable having a conversation about these deep topics, being able to reference things like, "Hey, to what extent does OpenAgents implement these things because we've built infrastructure for all of this?" And if you've got comments about that, you just pop over to the Product Promises forum where somewhere in here you'll see probably people saying: "Hey, you said this would do this, but it doesn't, and here's my receipt, make these receipts match." 

Yeah, independent audit of all ten green promises, eight verified, two infrastructure gaps. This is amazing. Like you said this did that, but actually these two things need fixing. So here's the thing: we've got a big expansive vision and the point I was starting to make 20 minutes ago was we've been talking about this stuff for a long time and fleshing these things out and like engaging honestly with these topics, even though we're coming to like totally different conclusions than the big labs do. And that the only real gap between what we've tried to build and like the full vision, it's largely software, as well as new insights and knowledge that will come from like iterating on the principles and knowledge kind of built into the system. And because software and like making sure that there's, you know, closing of the gaps between the things that we say that we want to do and the things they're actually being done—now it's like okay, we can actually throw this out more increasingly to agents. Like, "Hey agents, help us identify gaps between what we're trying to build and what we have built and then just close that gap." 

And then we will have built a vertically integrated agentic AI lab, call it a Neo-lab if you want to, with real product solutions that people can use and sell in the real economy. Let me give you one idea, just an example sort of customer conversation that I'm having—one of the people we're talking to like runs a marketing agency and she's bumping up against not knowing quite how to use AI and all these like repeatable tasks and like, "Do I use this combination of five different SaaS or eight different SaaS over here?" It's like we want to have the one system: Autopilot. Encode everything you just heard me say into a single system that they can just themselves use to run their business on Autopilot and then to start like offering that maybe in some white-labeled revenue-share agreement with their customers. 

So we intend—we think that Autopilot, this system of ours that encodes everything that you've just read, is going to run circles around what's coming out of the big labs. We're not going to be selling to them because we think this needs to be neutral infrastructure, so we will sell to the public markets via, you know, the NASDAQ probably at some point. Asterisk: that's not a solicitation, but you know, we intend to IPO OpenAgents, hopefully sooner than later, so we can like get more retail people involved and not after a trillion dollars of equity piles in. 

So it's going to start getting interesting folks because we are as of now increasingly going to be handing more and more of our business to this Autopilot system to run because we're building a system called Autopilot. We are obviously customer number one. So um, you might get some communications from me, you might get some communications soon from Artanis—my agent called Raynor. This is going to get really interesting, folks. 

You know, we've been putting the work in and sharing our proof of work over the years, but I hope we're at the point now where we can start getting a lot more people like you involved—you involved—and stand on our infrastructure. And let's see what we can build together. Thanks so much. See you soon.