Why the Agent Era Needs a New Operating System | DHH
- 01Linux's historic weaknesses are now its agent-era strengths
- 02A third desktop OS window has opened for the first time in 25 years
- 03The malleable OS: software that mutates on request
- 04Application lock-in collapses, which reopens mobile
- 05Agents eradicate entire application categories, not just replicate them
- 06The speed trajectory: live-mutating software at 1,500 tokens a second and beyond
1. Key Themes
Linux's historic weaknesses are now its agent-era strengths
DHH's core thesis is that the three traditional complaints about Linux are exactly what agents need. A terminal-first interface, text config files, and verbose errors give agents a native surface to operate on, while mouse-and-click operating systems don't.
"Every single thing that made Linux hard to use in the past, oh, it's very terminal heavy, it's got a bunch of config files that's hard to figure out, and it's got inscrutable error messages. These are the three things agents love." — David Heinemeier Hansson 00:00:00
He gives a concrete example of why macOS fails: "If you want to remap the hotkeys to go to a different workspace in macOS, there's only one way to do that. So manually go into a settings app, click one field at a time... You cannot do that agentically." 00:07:06 On Linux, "They want a command line interface for everything. They want predictable config files that they can just edit and update. They want all the details in the error messages." 00:07:32
A third desktop OS window has opened for the first time in 25 years
DHH argues that OS markets have been locked for decades by application availability, and that lock-in is dissolving.
"The window is now open for a third operating system... The last time we had an opening was that OS X 10 in 2001. That was the last time anyone premiered a new operating system that made any dent in the global market share. Since then, nothing. Window closed. This year, window opened." — David Heinemeier Hansson 00:08:30
His ambition is concrete: "Within, let's call it three years, we're going to be bigger than Mac OS. There is only 180 million Mac OS users in the world... that's not even a fucking C-tier internet service." 00:34:24
The malleable OS: software that mutates on request
Casado and DHH describe an OS where users change any element by prompting.
"They want a new start menu. They want to move the clock around. They want a different calendar. They just want to add a feature to an existing part of the operating system. And they can just do that by prompting. And it literally happens in sometimes seconds, maybe at most a few minutes. It's deeply intoxicating." — David Heinemeier Hansson 00:06:14
Casado confirms it from firsthand experience: "Just tell your agent, modify the OS on which I'm operating right now, and it does it and it works." 00:00:17
Application lock-in collapses, which reopens mobile
Since ecosystem moats were about app availability, near-zero software cost breaks them. DHH cites a community member who replicated Adobe Lightroom in two weeks.
"One guy over the course of two weeks replicated all of Lightroom. Not just replicated, but made it better. And Adobe had literally spent, I don't know, a decade and a half and tens of millions of dollars, if not more, developing." — David Heinemeier Hansson 00:10:00
Casado extends this to mobile: "The concept of building a third mobile ecosystem five years ago... was just, I mean, certain to fail. Completely daunting. And now for some token budget, we can replicate the top 10,000 apps. And for many apps, they may not even need to be replicated because the agents can just natively do the work." 00:11:07 DHH's motto: "First we take Manhattan, desktop. Then we take Berlin, mobile." 00:10:58
Agents eradicate entire application categories, not just replicate them
DHH says he no longer uses a word processor or Photoshop, because an agent handles the job directly.
"I just asked my agent to come up with the contract and then I will review it in Markdown. And when I'm happy, I'll just have it make a PDF... I don't even sign my documents anymore... The agent has my signature and it'll put it on the document and then it will mail it out." — David Heinemeier Hansson 00:11:31
On Photoshop: "Historically, especially something like Photoshop, was not just an application. It was a profession... agents have taken all that on and now they can do it without the app." 00:12:58
The speed trajectory: live-mutating software at 1,500 tokens a second and beyond
DHH points to inference speed as the next unlock, citing Fireworks AI's demonstration.
"Fireworks just two weeks ago at their big conference demonstrated 1,500 tokens a second... In four seconds, they went from the prompt, hey, can you make me a Tetris game to the game running... Everything just mutates live." — David Heinemeier Hansson 00:13:24
He extrapolates: "What's going to happen two years from now when these models are going to run at 100,000 tokens a second? A million tokens a second?" 00:00:26
Tiny teams versus giants, and the compression of the VC model
Omarchy's current version was built part-time in about two and a half months by DHH and Ryan Hughes, competing against thousands of Microsoft programmers.
"Now you look at venture capital, right? There's literally a, what, two-decade-long history where all the money just plowed into developers because you needed developers to build the thing. And you also needed a runway of several years... Now, suddenly all of that is getting compressed down to days, weeks, at worst months." — David Heinemeier Hansson 00:15:40
Human taste still matters: the last 2% and agent-driven perfectionism
Scott Kupor challenges that software moats live in long-tail design choices. DHH agrees and says agents amplify taste.
"I feel like I can get more of my taste, more of my vision, more of my perfectionism into the software because the barriers between me spotting something that's not quite right, a few pixels that are off and wanting to fix it has just collapsed." — David Heinemeier Hansson 00:17:40
He adds: "The most gratifying for some... is the first 80%... Now, for whatever reason, I am blessed with the entire dopamine payoff in the last 2%." 00:19:02
Open weights keep the frontier honest, and the fear of an AI duopoly
DHH wants both frontier labs and open-weight models, and warns against a Google/Meta-style value capture.
"I think it's absolutely essential that we have open source and we have open weights to keep the whole game honest. We need competition. I mean, I have spent two phases of my career being very dismayed by duopolies splitting everything up." — David Heinemeier Hansson 00:26:22
On local models: "There is intelligence in a box that you can fit on your damn desk that you can buy for single-digit thousands of dollars. That's incredible." 00:29:52 He's not an open-source purist: "I cheer for the fact that the U.S. have these two frontier labs and that they do have the lead because we get to then build with that stuff." 00:30:22
The return of the hacker, pseudonymous builders, and demo culture
DHH sees AI re-democratizing domains once gobbled up by big companies.
"The return of the pseudo-anonymous hacker... I love this energy around computers again like it was in the 90s, the 80s where hackers and experimenters are suddenly at the forefront of everything." — David Heinemeier Hansson 00:44:59
He frames X as the connective tissue: "X has placed such an enormous connecting tissue role for this whole revolution, for the artist, for the pseudo-anonymous hackers, for the local AI movement." 00:47:28
2. Contrarian Perspectives
Source code is not a moat, and anyone who thinks it is, is delusional
Most software investors still treat large proprietary codebases as defensible. DHH, who built Basecamp, Hey, and Rails, says that thesis is already dead.
"Anyone who still thinks that there is moat in source code, even millions of lines of it, or that that is going to protect their ecosystem are delusional." — David Heinemeier Hansson 00:16:09
His evidence is that a handful of people replicated Lightroom in two weeks and built a competitive OS in about 2.5 months.
Apple and Windows are "unfit" for agents, and Apple is the "AOL of the agent era"
Apple is widely seen as the safest consumer platform. DHH argues its walled-garden instincts will make it irrelevant to agent-native users.
"Apple is running the risk right now of becoming the AOL of the agent era. We're going to provide this wall garden... And most people don't want AOL. They want the open internet." — David Heinemeier Hansson 00:08:01
He also notes Apple is "making it even harder to use macOS with agents. Even more scary pop-up screams about how you shouldn't let Claude have access to your files." 00:05:20
Linux on the desktop wins, and Mac is a small market
The "year of Linux on the desktop" is a decades-old joke. DHH argues it is now plausible and that the Mac market is small.
"There is only 180 million Mac OS users in the world. The Mac is not that big. I know it looks like it in America, and especially if you walk around Silicon Valley where everyone uses a Mac. But outside of that, tiny." — David Heinemeier Hansson 00:34:24
He wouldn't "put majority odds" on unseating Mac in 12 months, "but I also wouldn't rule it out." 00:35:52
Individuals with agents beat 60,000-person companies, and you should avoid thinking about AI beyond the current model
DHH prefers a "fleet of agents" with zero coordination cost to scaling headcount: "Dealing with the productive potential of 60,000 people through agents. Now, now you have me interested, right? No fucking argument. No HR. No committee meetings." 00:20:24
He also rejects long-horizon forecasting and pessimistic safety thinking: "You try to imagine two, three, four model hops ahead, you're just going to go crazy... I think this is why so many of these AI safety people end up fucking crazy." 00:36:51 He compares it to Oppenheimer: "The bomb was the best thing that ever happened to world peace. We literally got 50 years of a beautiful Cold War that did not include total nuclear annihilation." 00:37:49
Local AI surprised even skeptics, and hardware may be where the "takeoff" stalls
DHH admits he was wrong that local models would be useless: "I wrote a piece, I don't know, six months ago... saying, like, this is stupid. Why are people buying fast personal computers? ... And then thankfully I was totally wrong." 00:29:24 He also observes a split on timelines: "The folks working on the software side of things are like, oh my God, this is just about to happen. And the people working on actuators and mechanics... are like, ah, I don't know. Maybe it's actually quite hard to peel an onion." 00:25:09
3. Companies Identified
Omarchy (Omachi in the transcript)
An agent-first Linux distribution created by DHH, with a nonprofit foundation behind it. Built with agents and designed for malleability, a terminal aesthetic, and 90s demo culture. It is the central subject of the episode.
"I'm doing a Linux distribution called Omachi, the first native OS for the age of agents... getting hundreds of thousands of people to switch from Windows and Mac to a Linux system." — David Heinemeier Hansson 00:03:06
Omarchy / Omicom Foundation
The nonprofit behind Omarchy, which sponsors contributors, including pseudonymous hackers and local AI builders.
"I've been setting up the Omicom Foundation, which is the nonprofit behind Omachi. We've raised, I don't know, $23 million in tokens and donations." — David Heinemeier Hansson 00:12:00
Fireworks AI
Inference provider. Why mentioned: it demonstrated 1,500 tokens per second code generation, previewing live-mutating software.
"Fireworks just two weeks ago at their big conference demonstrated 1,500 tokens a second... In four seconds, they went from the prompt, hey, can you make me a Tetris game to the game running." — David Heinemeier Hansson 00:13:24
Anthropic (Claude / Opus)
Frontier lab. DHH is a heavy user and credits rapid model gains.
"At the time I've been developing Omachi, we went from whatever, Opus 4.8... And now we got Opus 5.5. The distance between Opus 4.8 and 5.5, enormous." — David Heinemeier Hansson 00:16:09
On cost: "I got three clot subscriptions. That's it... Most weeks, I'm on one $200 20X sub and I could do everything." 00:21:44
OpenAI
Named as one of the two leading frontier labs. DHH says he appreciates them but fears a duopoly.
"As much as I appreciate OpenS5.5 and what OpenAI is doing, I think it would be a nightmare scenario if we end up with a duopoly position." — David Heinemeier Hansson 00:27:28
SpaceX / xAI and Google (competitive frontier labs)
Cited as evidence of healthy closed-model competition.
"The fact that we don't just have these two labs that are ahead, but we have SpaceX and AI quite close to it. We have Google that seemingly one moment is in the game and then the next moment somewhere off smoking something funny on an island somewhere." — David Heinemeier Hansson 00:28:24
Scott Kupor also notes that Grok's bot "keeps ordering home automation stuff." 00:25:28
Quickshell
The malleable shell used for all desktop interactions in Omarchy, created by the pseudonymous "outfoxed."
"Outfoxed, the creator of Quickshell, the malleable shell that we're using for all the desktop interactions." — David Heinemeier Hansson 00:44:59
Open-weight models: GLM, DeepSeek, Qwen
DHH points to open-weight models roughly six to seven months behind the frontier.
"GLM-53 or DeepSeek Flash, I don't know, six, seven months behind-ish on some tasks... QN38 Flash Next is a totally usable model." — David Heinemeier Hansson 00:29:24
NVIDIA DGX Spark
Desktop hardware capable of running useful local models: "A DGX Spark can run a useful model. You can actually use for something." 00:29:52
Unitree
Chinese robotics company. DHH visited and found their timelines for general intelligence far longer than Silicon Valley's.
"Unitree guys are like, ah, I don't know, four years, five years, maybe 10." — David Heinemeier Hansson 00:25:09
Basecamp / 37signals and Hey
DHH's own company (about 60 people). Basecamp is the original SaaS project management tool; Hey is the email service launched about five years ago.
"Basecamp. The original project management tools, SaaS before it was called SaaS, released the day before or the day after Facebook went live." — David Heinemeier Hansson 00:01:41
Ruby on Rails
Web framework extracted from Basecamp, used by Twitter, GitHub, Shopify, Airbnb, and Coinbase.
"Ruby on Rails was then used to build Twitter. It's been used to build GitHub, Shopify, Airbnb, Coinbase, and a million other applications out there." — David Heinemeier Hansson 00:02:11
Shopify
Referenced through Toby Lütke's workflow and as a Rails success story. It has roughly 7,000 employees per DHH's account.
Adobe (Lightroom, Photoshop)
Cited as a case of lock-in collapsing. A community member rebuilt Lightroom in two weeks. 00:10:00
Apple and Microsoft
Targets of the thesis. Apple faces "AOL of the agent era" risk; Microsoft failed in mobile because it lacked apps.
"Microsoft tried. They spent, I don't know, several billion dollars to make Windows OS a thing on mobile. And it wasn't a thing. In huge part because they did not have the applications." — David Heinemeier Hansson 00:09:30
Google and Meta
DHH's cautionary tale of duopoly value capture.
"Those two companies sucked out 99% of the value chain in online advertisement." — David Heinemeier Hansson 00:30:50
Mia's Lab and 0x0 (local AI projects)
Pseudonymous local AI projects sponsored through the Omarchy foundation.
"0x0 is someone we sponsor through the Yamakon Foundation. Mia's Lab, also both pseudo-anonymous." — David Heinemeier Hansson 00:45:58
X (Twitter)
Described as the connective tissue of the hacker, artist, and local AI revolution. 00:47:28
Other briefly named
- Ubuntu: DHH's first desktop Linux, about two and a half years ago. 00:03:35
- Windows Phone: cautionary tale of an ecosystem without apps. 00:09:30
- StarCraft: used by Kupor as an example of software moat through tiny design tweaks. 00:16:26
- LinkedIn: the "anti" culture DHH contrasts hackers against. 00:47:00
4. People Identified
David Heinemeier Hansson (DHH)
Creator of Ruby on Rails, co-founder of 37signals (Basecamp, Hey), author of four books, and now creator of Omarchy. Why mentioned: he is the guest and an exemplar of the agent-augmented solo builder.
"I've not written a single line of code for Omachi by hand, but I've written hundreds of thousands of lines of code through agents." — David Heinemeier Hansson 00:05:20
Ryan Hughes
DHH's main partner on Omarchy.
"Me and Ryan Hughes, my main partner on this project, that was the version that turned into this thing that's now taking off like crazy." — David Heinemeier Hansson 00:14:48
Toby Lütke
Shopify CEO, cited for the "hyperdrive" experience of working with agents on weekends and then returning to a large company.
"Toby Lutka calls this when he does it over the weekend and then has to go back into the office of a literate company of 7,000 people. Like coming out of hyperdrive." — David Heinemeier Hansson 00:20:46
John Carmack, Tim Sweeney, Mitchell Hashimoto
Legendary programmers that DHH says are fully "pilled" on AI. Scott Kupor adds Eric Raymond.
"As you mentioned, Carmack or Tim Sweeney or Mitchell Mashimoto or there's a million of these people I deeply respect... 100% of them are leaned in. They're pilled." — David Heinemeier Hansson 00:44:31
Eric Raymond
Open-source pioneer, mentioned by Scott Kupor as having returned to guide the AI wave. 00:39:17
Outfoxed
Pseudonymous creator of Quickshell, on the Omarchy core team. 00:44:59
Hancore
Pseudonymous artist on the Omarchy foundation staff list. 00:44:59
0x0
Pseudonymous local AI builder sponsored by the foundation. 00:45:58
Elon Musk
Cited both for ambition ("let's populate all the planets") and for rescuing X from "blue-haired bandits." 00:36:22
Jonathan Ive
Credited alongside Apple with "aluminum polish all character out of the computer." 00:51:00
Dwarkesh Patel
Referenced via a podcast interview in which a guest predicted superintelligence by next year. 00:25:09
Martin Casado, Scott Kupor, Anish Acharya
a16z hosts. Casado confirmed the malleable-OS experience firsthand; Kupor offered the "ASI test" of an AI building its own robot with a $20,000 budget.
"I gave it, like, a $20,000 budget and I'm like, just go build your robot... we're just not at the point where it's even close to being able to do that." — Scott Kupor 00:23:42
5. Operating Insights
Use agents to turn every "not quite right" into an immediate fix
The operating shift is that the gap between noticing a flaw and fixing it has collapsed, so teams that care about craft can ship at a higher quality bar.
"I now see any problem, any problem at all, any piece of software that is not perfect as an opportunity to go prompt until it's perfect." — David Heinemeier Hansson 00:17:40
Staff roles accordingly: the scarce resource is the person with the eye who won't stop at the first 80%. As DHH says, "You need the eye for a moment still." 00:19:55
Pay for subscriptions, not metered tokens, and you may not even hit limits
DHH reports running a frontier-level workflow on subscriptions alone.
"I got three clot subscriptions. That's it. And I can remember two weeks out of the last three months where I actually maxed or got close to maxing those out. Most weeks, I'm on one $200 20X sub and I could do everything." — David Heinemeier Hansson 00:21:44
The takeaway is that agent leverage is not capital-gated.
Replace coordination cost with a fleet of agents
His organizing principle: "Just me high bandwidth with a fleet of agents executing on a distilled vision." 00:20:46 He frames the unit economics of small teams as "no HR, no committee meetings," which argues for staying small and pushing agent leverage before adding headcount.
Dogfood the agent: build the product with agents and ship agents to users
Omarchy's inflection came from doing both: "We built it with agents, we put agents in the hands of everyone, and seeing hundreds of thousands of people pick it up." 00:35:23 Building with agents accelerates your own roadmap, while exposing agents to users makes the product itself the platform.
Pursue narrow, ferocious ambition and avoid long-horizon speculation
"I try to be ferociously ambitious within this narrow scope and just put all the energy into that. And then also just not try to think about what 12 months from now look like." — David Heinemeier Hansson 00:36:22
His rationale is mental-health and focus: stay in "AI euphoria" rather than "AI psychosis" by anchoring to what the tools do today.
6. Overlooked Insights
Agent-driven security prompts on macOS are a strategic tell about platform incentives
DHH mentions in passing that "there was just another drop today that Apple is making it even harder to use macOS with agents. Even more scary pop-up screams about how you shouldn't let Claude have access to your files. Well, what is it going to do then if it doesn't have access to your files?" 00:05:20 This reveals that platform owners are structurally incentivized to restrict third-party agents, since agents bypass the app-store toll booth. It suggests the platform conflict of the agent era (who controls the file system and OS-level actions) is already beginning, and that open platforms with fully permissive agent access may win developer and power-user mindshare.
Mobile's duopoly may be weaker than it looks because the 30% toll depends on app scarcity
In a brief aside, DHH ties the entire mobile duopoly to application availability: "The entire lock-in on these ecosystems is about application availability. And in the age of agents, software development has just dropped to zero in terms of pricing, in terms of difficulty." 00:10:30 Combined with Casado's note that agents can natively do the work an app was signed up to do, this implies the real threat to the app-store tax is not an alternative store but the disappearance of discrete apps. For investors, it points to a window in which mobile-adjacent bets, from agent-native phone OSes to app-replacing agent layers, may be undervalued relative to desktop.