DHH on Decoded Genius: How AI Agents Flipped Him in Six Months, Why Grinding Is for Computers, the $2 Million Cloud Exit, Jeff Bezos's Stupid Number, and the Stoic Case for Optimism

David Heinemeier Hansson, the creator of Ruby on Rails and co-owner of 37signals (Basecamp, HEY), sat down with Decoded Genius for a two-hour conversation that starts with how AI coding agents changed his mind in a matter of weeks and ends with Stoicism, yellow Lamborghinis and why supercars are turning into mechanical watches. In between: why he never took venture capital, the one-page term sheet Jeff Bezos accepted, the math behind leaving the cloud, the ONCE experiment that failed, the HEY fight with Apple, and why the 80-hour founder week is a delusion.

TLDW

DHH went from writing 37signals’ new kanban tool Fizzy almost entirely by hand to starting every piece of code with a prompt, after agent harnesses and a new wave of models arrived around December and made the agents’ code good enough to merge. The key change was the feedback loop: minutes instead of weeks, which makes agents feel like a bionic arm rather than a team to project-manage. He argues constraints matter more than ever now that implementation is cheap, that software entrepreneurship is open to people with ideas and a $20 Claude subscription but is 100x more competitive, and that distribution plus product is now table stakes. On money: bootstrapping came from dot-com bust trauma, Bezos bought secondaries with no control at a price 37signals set to make him go away, and raising for a pure software play today is one of the worst times in history because nobody can extrapolate SaaS revenue past the AI joker. He defends progress with the ATM, tractor, Model T and nuclear power examples, pushes back on AI doom by comparing it to earlier doomsday stories, and leans on the Stoic split between what you control and what you do not. On work: a 40-hour moving average, the power law of hours, the “lifestyle business” insult as a compliment, the disappointment of hitting your number, and why independence is priceless. On operations: the cloud exit cut roughly $3.2 million a year to about $800,000 with the same team, ONCE (one-time-fee software) “didn’t work yet,” and the Apple fight over HEY turned into a free launch campaign worth 30,000 paying customers in under two weeks. It closes on cars, the manual-transmission revival as humans wanting to stay in the loop, and one piece of advice: read A Guide to the Good Life.

Thoughts

The most useful thing DHH says about AI coding is not that the models got smarter. It is that the feedback loop collapsed, and that the collapse changed what kind of work it is. He expected supervising agents to feel like managing human programmers, which he dislikes, and it turned out to be nothing like it because you review the result in seven minutes instead of two weeks. That is a sharper explanation of the “vibe shift” among senior engineers than most of the benchmark talk. It also explains why his reservation was aesthetic rather than practical: once the code looked like his code with a little nudging, the objection disappeared. Taste was the gate, and taste got cleared.

The constraint argument early in the conversation is the one founders should tape to the monitor. If implementation stops being the bottleneck, every idea you have becomes buildable, and that is a curse, not a gift. DHH says he would not trust himself with a team of 100 because he would pursue 50 ideas instead of the essential 10. Agents give everyone that team of 100. The small-team discipline that made Basecamp coherent was partly a product of scarcity, and it now has to be chosen deliberately. Expect a lot of bloated, incoherent software from people who confuse capacity with direction.

The cloud exit segment, about 85 minutes in, has the cleanest piece of business math in the interview and it applies well beyond servers. A dollar saved goes straight to the bottom line; a dollar of new revenue at 20% margins leaves you twenty cents. So cutting a million in costs is worth roughly five million in new sales, and it was far less work. His point about why managers do not do this is the underpriced part: professional managers are paid on the multiple, and the multiple is driven by revenue and growth, so nobody is rewarded for the boring cut. Owners are. The fact that 37signals did not hire or fire a single operations person on either side of the move is the detail that undercuts the whole “the cloud saves headcount” pitch.

The ONCE story is worth more than it looks because DHH tells it honestly: the generic-drugs thesis for mature software categories did not work, and he insists on saying “yet.” His evidence is his own history. 37signals had a chat product called Campfire years before Slack and the market did not want it; Slack arrived later with the right timing and built a huge category. Being too early looks identical to being wrong until someone else is right. The companion insight, that Basecamp’s private LLC numbers kept competitors away for a decade because everyone underestimated the business by one or two orders of magnitude, is a reminder that secrecy about how good a niche is can itself be a moat.

The best moment comes in the car talk near the end, where it quietly ties back to AI. Supercars cannot win on acceleration against a Model S Plaid, so they will become mechanical watches: worse at the measurable job, loved for the feel. The manual-transmission revival happened because people want a human in the loop, and DHH connects that directly to agents, saying that as more of life stops needing us, we look for things that still do. Put that next to his earlier claim that most people need work and would be depressed, not artistic, in an AI abundance utopia, and you get a more honest picture of the transition than either the doomers or the boosters usually give. The economic case for progress can be right and the need to be needed can still go unmet.

Key Takeaways

As late as early December, 37signals had written virtually all of its new kanban product, Fizzy, by hand, and DHH told an interviewer AI was not good enough yet.
He disliked the autocomplete era of AI coding, comparing it to someone constantly interrupting you mid-sentence.
Around December, new models and agent harnesses (OpenCode, Pi, Codex, Claude Code) moved coding from autocomplete to missions, and he flipped 180 degrees.
He now starts every new piece of code with a prompt and only occasionally drops in to write code himself.
Nobody, even in the AI community, predicted the terminal would become the main agent platform, which is why he urges humility about even six-month forecasts.
Agents do not feel like managing programmers because the review loop is minutes, not weeks or months.
His real objection was aesthetic: early agents produced working but ugly code. Once the code looked like his, he embraced it as a bionic arm.
Programmers who loved the mechanics of turning requirements into code, and not the outcomes, are struggling most, and he sympathizes with that sense of loss.
He learned to program because he wanted programs, and agents have returned him to that original motivation.
Bootstrapping came from watching the dot-com boom and bust; he compares it to Depression-era parents who always offer extra butter.
He built Basecamp’s technical side alone with 20-year-old tools, while Jason Fried and Ryan did design, which made him skeptical that anyone needed 100 people.
At one point 37signals had seven people and four major products and launched a new product every year.
Constraints force you to kill your darlings; unlimited implementation capacity through agents will produce bloated software for those who do not choose limits.
The “idea guy” with no money and no implementation skill can now validate a prototype with a $20 a month Claude subscription.
The flip side is a market 100x more competitive than 20 years ago, where even pedigreed teams launch to a blip and then crickets.
To break through you now need both great distribution and a great product; bad distribution with okay software is a closed window.
37signals turned away around 40 VCs, then met Jeff Bezos because his reputation (itself a form of distribution) made the conversation interesting.
Bezos bought secondary shares: money to the founders, zero control, no timeline, no guaranteed exit, at a one-page “stupid number” meant to make him go away.
Taking money off the table made the founders immune to later lavish VC offers.
Raise capital when you need to build factories or inventory, not to pay salaries to find out whether an idea has legs.
The SaaS valuation music has stopped: nobody will make 15-year recurring revenue projections when no one knows what the world looks like in six months.
He has made peace with every outcome for Basecamp, including AGI eating it this year, using Stoic negative visualization.
Productivity gains are individually tragic but necessary for prosperity; the ATM increased bank-teller jobs by making branches cheaper to open.
He blames much current political discontent on the glib elite response to deindustrialization: a cheaper iPhone does not replace a job.
No single nation gets to pause progress; if one country stops AI, others will not.
Pushed on AI existential risk, he argues it requires more extrapolation than the nuclear threat did and compares it to earlier doomsday narratives.
The West’s retreat from nuclear power is his example of what slowing a technology costs; France built most of its generation capacity in the 1980s.
Society accepts about 42,000 US traffic deaths a year as a tradeoff; AI deserves the same tradeoff thinking, not a return to the box.
The 80 to 120 hour founder week is a delusion for almost everyone; his 25-year moving average is 40 hours.
Hours follow a power law: his 10 hours a week on early Basecamp were the most productive of his life.
“Lifestyle business” is a compliment: a lifestyle and a business at the same time.
Most people need work for meaning; early retirees and founders who sold too early are often the most miserable people he knows.
Hitting your number does not fix your head, and nobody believes that until it happens to them.
Independence, including freedom from public markets, is something you cannot price until you lose it.
Passion can follow effort; things become more fun as you become competent, including expense reports.
Cloud spend had become 37signals’ second largest expense after payroll; leaving cut it from about $3.2 million to about $800,000 a year with the same team.
A dollar saved goes fully to the bottom line, while a dollar of revenue at 20% margins leaves twenty cents; 37signals shares 10% of profit with employees.
ONCE, selling commodity software such as Campfire for a one-time fee, did not work, and he insists on saying it did not work yet.
Private LLC numbers led people to underestimate Basecamp by one to two orders of magnitude, which kept competitors away for about a decade.
Apple’s attempt to take 30% of HEY in 2020 became a free launch campaign he values at $20 million, bringing 30,000 paying customers in under two weeks.
He counts the Apple fight and the 2021 no-politics-at-work policy, after which about 30% of staff left, as his two proudest career moments.
Supercars will become mechanical watches, and the manual-transmission revival reflects people wanting to stay in the loop.

Detailed Summary

From hand-written code to agents in six months

DHH opens by admitting how recently he changed his mind. At the start of December, 37signals had built its new kanban tool Fizzy almost entirely by hand, and he told people who were already “fully AI pilled” that the tools were not good enough. He hated the autocomplete model of AI coding. Then a major model release landed alongside agent harnesses like OpenCode, Pi, Codex and Claude Code, and the interaction changed from guessing your next line to taking on a mission. The code coming back was code he wanted to merge. Barely six months later, every new piece of code he writes starts as a prompt. The unexpected return of the terminal as the center of agent work, now used by non-technical people too, is his evidence that nobody can forecast even the end of the year.

He thought directing agents would feel like project-managing human programmers. It does not, because the feedback loop is minutes rather than weeks. Once the agents’ output looked like his own code, the experience felt like gaining a bionic arm, not giving work away to smarter robots. He acknowledges humans may become the dumb element in every loop within two years, but for now he calls this the most exciting time to be a programmer. The people struggling are the ones who loved the mechanics of coding more than the outcomes, and who were well paid for being the constrained resource.

Why 37signals stayed bootstrapped

Watching the dot-com bust marked DHH and Jason Fried the way the Depression marked an earlier generation. They saw overfunded companies collapse into layoffs and wanted to grow slower on expenses while keeping big ambitions. In software, you did not need to build factories. DHH wrote Basecamp’s technical side alone, and at one point the company ran four major products with seven people. He considers large teams actively harmful, producing bloated software, and says he would not trust himself with 100 people because he would chase 50 ideas instead of the essential 10. Constraints act as a funnel that yields coherent products, and he uses the original Star Wars against its CGI-heavy successors as the example.

Starting a company now: easier and 100x more competitive

He once wrote that there is no room for the idea guy, because building required either skill or money. That has changed: someone with a $20 Claude subscription can build a prototype and validate an idea. The monkey’s paw is competition. When everyone can build, ideas need defensible angles and distribution matters more than ever, because the days of organic discovery for merely nice software are gone. His answer to the host: you now need both great distribution and a great product, and if you must pick one, distribution gives you better odds of a moderate success.

Jeff Bezos and the one-page term sheet

After Basecamp’s launch roughly 40 VCs called and were turned away. Bezos was different because his reputation made the conversation worth having. 37signals did not need money, so the founders wrote a single page with three key metrics and a deliberately ludicrous number for a secondary sale: millions to the founders, no control, no timeline, no guaranteed exit. Bezos’s team called it ridiculous; Bezos, who had read Getting Real, took the bet anyway. Nearly 20 years later his stake is worth many times what he paid. DHH describes it as “rational irrationality” and credits it with making him immune to later VC offers. The relationship amounted to a yearly conversation where Bezos mostly told them they already knew what to do.

Raise or not, and the end of the SaaS machine

His rule: raise if the business needs capital for factories or inventory, be skeptical if the money only pays salaries, especially now that agents make early validation cheap. He argues pure software valuations peaked around 2021, when LTV to CAC ratios made SaaS look like a machine you could extrapolate for 15 years. The AI joker has ended that, so raising for software now means poor multiples and hard defensibility questions. He sees a blessing in it, since the lack of easy money forces founders onto the path 37signals took from Chicago.

Progress, jobs and AI doom

DHH says Basecamp, launched in February 2004, has already given him everything he could expect, so he has made peace with it ending. He applies the Stoic split between what he controls (building, showing up, earning attention) and what he does not (whether AGI sweeps away SaaS). Smashing data centers is the modern Luddite response, and he calls job loss individually tragic but societally necessary, citing the ATM, the tractor and the Model T. He does take political discontent seriously, blaming the glib elite response to deindustrialization, and says a stable society needs most people to see things improving. But no single nation can pause progress, and the West’s abandonment of nuclear power shows what slowing down costs.

The host pushes back with the existential-risk argument, citing Roman Yampolskiy and the worry that a handful of labs could build something capable of killing humanity. DHH answers that humanity has always had a doomsday story, from religious judgment days to the nuclear standoff and climate predictions, and that AI risk requires more extrapolation than a single finger on a nuclear button did. He accepts that AI labs should take it seriously, but frames the rest as a tradeoff, like the 42,000 annual US traffic deaths society accepts for mobility. Being miserable about a future you do not control, he argues, makes you useless.

Hours, the lifestyle business and enough

Grinding is for computers, he says; you can run Claude all night as long as you pay the token bill. Humans need sleep and walks where the ideas that skip two months of drudgery arrive. His 25-year moving average is 40 hours a week, with a handful of 80-hour weeks. Hours follow a power law, and his 10 weekly hours on early Basecamp, squeezed between school and clients, were the most productive of his life. He rejects the American idea that you must choose between business and family, health and hobbies, and he takes the “lifestyle business” label as a compliment. Most people need work to feel good, so a post-work AI utopia would leave many people depressed rather than painting.

On money, he says a bootstrapped $10 million software business already puts you among the most fortunate humans who have ever lived. Your number depends on the lifestyle you want, and a private jet changes the math. Early retirement is often miserable, and founders who sold early learn that another great idea is hard to find. He expects Basecamp to be the best business idea he ever has. When the Bezos money hit his account he expected life to become magical, and it did not; he adds that no one believes this until it happens to them. What the money did buy was independence, which he would not trade for a bigger business with a boss, even if the boss is the public markets.

Passion follows competence

At 18 he wanted to make video games and would have found productivity software boring. He ended up loving Ruby and collaboration software, and even expense reports. Passion often follows effort, and most domains become interesting if you dig deep enough. He thinks you need to get excited about something in the business to last the seven to 10 years most companies need, whether that is the product, happy customers or taking care of employees. Lottery winners who quit work and burn out are his warning against wishing work away.

Leaving the cloud: rent or buy

The cloud question is rent versus buy. If you need something every day for years, you buy it, like a car you need 340 days a year. Going through expense reports, DHH found cloud spend had become the second largest line after payroll and was growing faster than the business. He calls the belief that running your own computers is exotic a marketing campaign by the hyperscalers, since nearly every internet company before about 2015 did it. The cloud never let them cut a single operations person, and the team stayed the same size after leaving. Spend fell from about $3.2 million to about $800,000 a year. Professional managers rarely do this because they are paid on the multiple, driven by revenue growth, not on profit. The host adds his own story of a garage server that beat a $200 a month cloud bill, and DHH says he is building tooling to bring back the closet computer.

ONCE, Campfire and being too early

ONCE was built on a generic-drugs thesis: mature software categories like team chat should have commodity prices, so 37signals sold Campfire, essentially Slack in a box, for a one-time fee. It did not work. Customers did not want to run their own machines, and SaaS convenience is real. He tells the story humbly: 37signals had Campfire as a SaaS chat product years before Slack, when the market still wanted the phone and conference rooms. Basecamp, by contrast, arrived at the right time, and private numbers let it enjoy the market almost alone for a decade. His lesson is that “we tried that, it didn’t work” is dangerous, because timing changes everything.

The Apple fight and the politics policy

In 2020, Apple approved the HEY email app and then reversed course days later, demanding in-app signup and a 30% cut. The fight played out on Twitter two weeks before WWDC, a David versus Goliath story millions of developers recognized. It ended in a face-saving truce, no 30%, and over 30,000 paying customers at about $100 a year in under two weeks, instantly a $3 million a year business. The risk was total, since roughly 80% of HEY’s buyers used Apple devices. A year later, 37signals banned societal politics at work, and a large share of staff left under intense public scrutiny. He calls both his proudest moments, justified by the same logic: he did not want to run the company that would come out the other side of backing down, and having made his money gave him the freedom to say no.

Cars, watches and staying in the loop

The week the Bezos wire arrived, DHH bought a yellow Lamborghini Gallardo. He got his license at 25 in Copenhagen, found racing at a track near Chicago, reached the 24 Hours of Le Mans grid in 2012 and won his class two years later. His favorite car is the Pagani Zonda, which he commissioned in 2010 after seeing it on Top Gear. He defends openly celebrating success as one of the best things about America. On electrification, he loves his new seven-seat Model Y, but argues supercars cannot win on acceleration anymore and will become mechanical watches, valued for feel rather than numbers. Heavy hybrids are stuck in the middle. The manual-transmission revival shows people want to be needed, and as agents need us less, he expects people to look for worth in things like a watch you wind or a clutch you work.

The one piece of advice

His final advice for founders is to read A Guide to the Good Life, the introductory text to Stoicism he credits for much of the worldview in the interview: negative visualization, separating what you control from what you do not, and building mental fortitude. He estimates 97% of his past stress was wasted, and says entrepreneurs, who deliberately put themselves in high-stakes situations, benefit most.

Notable Quotes

“If I can have this super fast loop, it’s more like an extension of myself. Like I get a bionic arm.”
David Heinemeier Hansson, on why coding agents do not feel like managing programmers

“You don’t want unlimited time and resources. You want constraints.”
DHH, on the danger of everyone suddenly having huge agent-powered capacity

“That music has stopped. No one is making 15-year projections on what kind of recurrent revenue a SaaS company can get today.”
DHH, on why this is one of the worst times to raise money for pure software

“Grinding is for computers. They don’t care. You can run Claude all night as long as you pay the token bill.”
DHH, on why founders do not need 80-hour weeks

“Everything is a goddamn power law. The first 10 hours you spend on your business every week matter infinitely more than whatever the 10 hours from 80 to 90.”
DHH, on his 40-hour moving average

“I have a lifestyle and I have a business. That sounds great. I get to have both at the same time.”
DHH, on the “lifestyle business” insult

“Independence is something you have a hard time pricing until you lose it.”
DHH, on why he would not trade 37signals for a bigger company with a boss or public shareholders

“We couldn’t even get rid of one person on our team going to the cloud.”
DHH, on the missing efficiency gains that led to the 37signals cloud exit

“What it did give us was essentially a $20 million launch campaign.”
DHH, on the 2020 fight with Apple over HEY

“I think supercars are going to turn into mechanical watches.”
DHH, on what combustion supercars have left once electric cars win on acceleration

Watch the full conversation with David Heinemeier Hansson on Decoded Genius here.

DHH on Decoded Genius: How AI Agents Flipped Him in Six Months, Why Grinding Is for Computers, the $2 Million Cloud Exit, Jeff Bezos’s Stupid Number, and the Stoic Case for Optimism