Tech Disruptions

DHH's AI Shift Why Agents Changed Everything

March 13, 202614:56Tech Disruptions

This episode explores 37signals' dramatic shift towards embracing AI, highlighting the transformative impact of "agent mode" where AI independently executes tasks rather than merely offering auto-completion. Listeners will learn how this approach significantly boosts developer productivity and how 37signals applies AI to reduce "toil" in areas like security report review, console compliance, and system diagnostics, effectively augmenting human work.

Key Takeaways

Detailed Report

37signals, a company known for its deliberate approach to technology, has undergone a significant shift in its stance on artificial intelligence. This change, particularly championed by CTO David Heinemeier Hansson (DHH), centers on the emergence of 'agent mode' AI, which has transformed their internal operations and is influencing their product strategy.

The Shift to Agent Mode

Historically, AI in development often manifested as intrusive auto-completion, a 'co-pilot' that DHH found disruptive. The pivotal change, occurring in late 2025 according to DHH, was the rise of 'agent mode.' This paradigm moves AI out of the text editor and into the terminal, where it operates more autonomously.

Instead of merely suggesting code, an AI agent is given a task and a plan, then uses the computer's tools—running bash commands, searching the web, or executing code—to achieve the goal. DHH likens this to hiring a junior developer: you assign a task, and they go off to complete it, often delivering an 80% complete solution that requires only minor human refinement. This represents a massive leap in productivity.

Automating Internal Toil

37signals has found practical, enterprise-level applications for these AI agents by targeting 'toil'—tedious, repetitive, but important programming work. Two key areas stand out:

Security Report Review

Reviewing security reports from platforms like HackerOne is a time-consuming necessity. The vast majority of these reports are low-quality, yet the critical few could prevent significant financial losses. 37signals now uses AI agents to pre-process these reports. The AI sifts through the noise, filters out low-value submissions, and allows human developers to focus their expertise on genuinely critical vulnerabilities. This acts as an advanced spam filter for security findings.

Console Access Reviews

Programmers at 37signals sometimes require access to production systems and customer data. Strict rules govern this access, necessitating manual review of all access logs for compliance. This is a monotonous task, often confirming that 'everything was fine.' AI agents, with their 'perfect patience,' are now trained to flag any non-compliant activity, freeing human time from this repetitive chore.

AI also assists with on-call duties and performance issues. By granting agents access to logs, exception systems, and monitoring tools, they can rapidly diagnose system degradations with impressive accuracy.

Product Strategy: User Agents and Democratized Creation

Despite internal success, 37signals has been cautious about 'jamming AI into everything' in their products like Basecamp or HEY. Early explorations into features like enhanced search or auto-completion didn't feel like a 'slam dunk,' risking 'AI slob'—adding features without unequivocal improvement.

DHH's 'Peak Experience'

A personal experience solidified DHH's optimism for AI's product potential. Faced with an obscure bug in Rails, he tasked Claude's Opus model with debugging it. The AI, acting like a seasoned developer, hypothesized, tried solutions, and even when initial attempts failed, persisted. It ultimately pinpointed the issue, found the problematic commit, and provided a working patch. This 'mind-blown' moment, which DHH compared to early technological wonders, demonstrated AI's power to solve complex problems that would typically take hours of human effort.

While not every AI interaction is perfect, DHH notes that the ratio of success is rapidly improving. Tools like OpenCode, which generate multiple working solutions from different models for a single prompt, further illustrate this progress.

Democratizing Software Creation

CEO Jason Fried highlights AI's role in democratizing software creation. While expert programmers gain speed, non-programmers can now build functional applications without deep technical understanding. Fried draws parallels to tools like Excel and FileMaker Pro, which empowered non-programmers to create solutions. However, he cautions that AI-generated code might be insecure or low-quality for critical systems, emphasizing the continued need for human oversight. For early-stage or less critical projects, the value of *any* working program created by a non-expert often outweighs the risks.

The Future: User-Supplied Agents

37signals' product strategy is nuanced. Instead of rushing to build every AI feature, they anticipate a future where users bring *their own* AI agents to interact with products. An agent might sign up for a Basecamp account like a normal user, learn about projects, and operate within the platform. This allows 37signals to focus on making their products simpler, clearer, and easier for external agents to integrate, for example, through command-line interfaces. This 'wait and see' approach with an open door aims to avoid investing heavily in features that might be quickly superseded by evolving AI capabilities.

AI, Layoffs, and Human Connection

Regarding the broader impact of AI on the workforce, Jason Fried holds a somewhat contrarian view. He suggests that many companies are often overstaffed, and AI might serve as a convenient justification for efficiency, rather than being the sole cause of layoffs. For 37signals, a lean company of about 60 people, AI is seen as a tool to develop faster with *fewer* people, not to eliminate roles.

In customer service, while AI can handle simple queries, 37signals prioritizes human connection. They maintain a highly trained, long-term human support team, recognizing that for nuanced issues or frustrated customers, human interaction is a 'massive competitive advantage.' They ensure customers always have the option to reach a human, preventing the common frustration of being stuck in an AI loop.

Ultimately, 37signals' journey with AI underscores several key insights: focus AI on specific 'toil' or friction points, embrace the autonomous capabilities of 'agent mode,' consider a strategy where users bring their own AI agents to your products, and recognize AI's power to democratize creation while maintaining crucial human oversight and connection.

Show Notes

DHH's AI Shift Why Agents Changed Everything

Source Materials

References & Resources

  • 37signals: A software company known for products like Basecamp and HEY, and for its deliberate approach to technology and business. The episode features insights from their CTO, David Heinemeier Hansson (DHH), and CEO, Jason Fried.
  • David Heinemeier Hansson (DHH): Co-founder and CTO of 37signals, creator of Ruby on Rails. His shift in perspective on AI, particularly "agent mode," is a central theme of the episode.
  • Jason Fried: Co-founder and CEO of 37signals. He discusses the broader implications of AI for product strategy, company size, and customer service.
  • HackerOne: A platform where external security researchers submit bug findings to companies; used by 37signals for their bug bounty program.
  • Basecamp: 37signals' project management and team communication software. Discussed in the context of AI integration and user-brought agents.
  • HEY: 37signals' email service. Mentioned alongside Basecamp as a product where AI integration is being carefully considered.
  • Ruby on Rails: A popular open-source web application framework created by DHH. An obscure bug in Rails was the subject of DHH's "peak experience" with an AI agent.
  • Claude 3 Opus: Anthropic's latest and most capable AI model, used by DHH in his "peak experience" to debug a complex Rails issue.
  • OpenCode: A tool mentioned by DHH that allows generating multiple code drafts from various AI models for complex tasks.
  • FileMaker Pro: A relational database application platform mentioned by Jason Fried as an example of a tool that empowered non-programmers to create applications.
  • Fizzy: 37signals' internal Kanban tool, used for early explorations into AI features like summarization.
  • Yie Ar Kung Fu: A classic fighting video game, mentioned by DHH to describe his "mind-blown" experience with early technology.
  • Commodore 64: A popular 8-bit home computer from the 1980s, referenced by DHH in his nostalgic comparison of technological wonder.
  • Netscape Navigator: An early and influential web browser, also referenced by DHH to describe a foundational moment of technological magic.

Glossary

  • AI (Artificial Intelligence): The simulation of human intelligence processes by machines, especially computer systems. These processes include learning, reasoning, and self-correction.
  • Agent Mode: A paradigm shift in AI interaction where the AI operates more autonomously, taking a task, devising a plan, and executing it using various tools (like bash commands, web searches, or running code) within a computer's environment, rather than just assisting with text input.
  • Auto-completion: A feature in software that predicts the rest of a word or phrase a user is typing, often used in text editors or search bars.
  • Bash commands: Commands entered into a Unix-like operating system's command-line interpreter (shell) to perform various tasks, such as navigating directories, running programs, or managing files.
  • Bug bounty programs: Programs offered by many websites and software developers by which individuals can receive recognition and compensation for reporting bugs, especially those pertaining to exploits and vulnerabilities.
  • CLI (Command-Line Interface): A text-based user interface used to run programs, manage computer files, and interact with the computer.
  • Commit: In version control systems (like Git), a "commit" is a snapshot of changes made to a project's files at a specific point in time, along with a message describing those changes.
  • Console reviews: A process at 37signals involving the manual review of access logs for production systems to ensure compliance with strict rules regarding data access.
  • Democratizing access: Making something (like software creation) available and understandable to a wider range of people, especially those without specialized skills or training.
  • Exception systems: Software systems designed to catch and log errors or unexpected events (exceptions) that occur during a program's execution.
  • Innovator's dilemma: A term coined by Clayton Christensen, describing how successful companies can fail by focusing too much on current customer needs and failing to adopt disruptive technologies or business models. In the podcast, it refers to new, imperfect AI tools eventually becoming robust.
  • Kanban tool: A visual system for managing work as it moves through a process, often used in software development to track tasks and workflow.
  • Logs: Records of events that occur in a computer system, used for monitoring, troubleshooting, and auditing.
  • On-call duties: The responsibility of being available outside of normal working hours to respond to urgent system issues or emergencies.
  • Patch: A set of changes to a computer program or its supporting data designed to update, fix, or improve it.
  • Performance monitoring tools: Software used to observe and analyze the performance of computer systems, applications, or networks, often tracking metrics like CPU usage, memory, and response times.
  • Production systems: Live, operational computer systems that are used by end-users or customers, as opposed to development or testing environments.
  • Rails: Short for Ruby on Rails, an open-source web application framework written in Ruby.
  • Terminal: A text-based interface used to interact with a computer's operating system, often used by

Sources / References

Full Transcript

HostOkay, so I thought I was getting a handle on the AI conversation, right? Been following it for years. But then you see a company, especially one as deliberate as 37signals, do a complete 180 on their stance, not just on using AI, but actively championing it. What changed so dramatically, and so fast, for them?
ExpertIt's fascinating, right? For David Heinemeier Hansson, their CTO, the shift wasn't just incremental model improvements, though those are happening. He pinpoints the latter half of 2025 as the real turning point, specifically with the rise of what they call "agent mode."
HostAgent mode. Sounds like something out of a spy movie. What's that actually mean for, say, a developer?
ExpertExactly! It sounds buzzwordy, and DHH even admits that. But the practical difference is profound. Before, AI in development was largely about auto-completion – like a really smart, but sometimes annoying, co-pilot trying to finish your sentences in your text editor. DHH found that disruptive. He wanted to type his own thoughts.
HostYeah, that's fair. I can relate. When I'm trying to write, I don't want someone constantly chiming in, "Did you mean 'paradigm shift'?"
ExpertRight. But agent mode is different. It's the AI moving *out* of your editor and into the terminal, running almost independently. You give it a task, a plan, and it goes off, using your computer's tools – running bash commands, looking things up on the web, even running code – to achieve that goal. It's not trying to complete *your* thought; it's executing *its own* plan you set for it.
HostSo, instead of being a backseat driver, it's like you've hired a junior developer who you give a task to, and they go off and do it?
ExpertPrecisely! And that changes everything. DHH describes it as going from a text input box where you'd ask for information or a second opinion, to something that's actually doing tangible work, where the output is often 80% there, and you just need to tweak the last 20%. That's a massive leap in productivity.
HostThat's a pretty compelling argument for any developer who's been skeptical. But it’s one thing to say it's useful for a developer in their personal workflow. How is 37signals, as an organization, putting these agents to work? Are they finding practical, enterprise-level applications?
ExpertAbsolutely, and this is where it gets really interesting for businesses. They've identified what they call "toil" – tedious, repetitive but important programming work that human developers often find draining. One significant area is reviewing security reports.
HostSecurity reports? Like, from bug bounty programs?
ExpertExactly. They use HackerOne, where external researchers submit findings. Now, the vast majority of these reports, according to DHH, are low-quality, even "bullshit." But you have to review them all, because that one in ten that's genuinely catastrophic, could save you millions. Reviewing these manually is incredibly time-consuming.
HostAnd probably soul-crushing. Like sifting through a mountain of spam emails to find one legitimate one.
ExpertPretty much. So, Jeremy, one of their team members, pioneered a system using AI agents. They give the AI access to historical reports, data on the submitters' past quality, and let it pre-cook the review process. The AI sifts through the sludge, filters the noise from the signal, and then human developers can focus their time on the high-value, critical reports. It's essentially advanced spam detection for security vulnerabilities.
HostThat's brilliant. It's not replacing the human, but augmenting them to do higher-value work. Where else are they applying this "toil reduction"?
ExpertAnother example is their bi-weekly "console reviews." Their programmers sometimes need to access production systems and customer data for investigations. They have strict rules about what can be accessed and why. So, they manually review all access logs to ensure compliance.
HostI imagine that's a lot of "everything was fine, everything was within jurisdiction" kind of paperwork.
ExpertExactly. It's tedious, boring work, especially when you're repeatedly finding nothing. DHH points out that AI is "perfectly patient" to do the same work over and over again. They can train it to flag anything that's not quite right, freeing up human time.
HostAnd for those moments when things *aren't* fine, or when a system goes down, can AI help there too?
ExpertAbsolutely. They've had success assisting with on-call duties and performance issues. When there's a system degradation, they can give the AI access to logs, exception systems, and performance monitoring tools. The AI can pull all that data, run queries, and often come up with a diagnosis that's "damn impressive" in DHH's words, and incredibly fast.
HostSo, these are all internal, behind-the-scenes applications. It sounds like they're building a highly efficient, AI-augmented internal operational team. But what about for their actual products, like Basecamp or HEY? They've famously been cautious about "jamming AI into everything." Has this internal success changed their product-facing strategy?
ExpertThis is the crux of it. For a long time, their early explorations into product features using AI – things like enhanced search or auto-completion within the product – didn't feel like a "slam dunk." They feared it would feel like "AI slob," just adding a sticker without unequivocally making the product better.
HostAnd they're not interested in features for the sake of being "AI-enabled." It has to be genuinely useful.
ExpertPrecisely. But this newfound enthusiasm from internal applications is pushing them to find areas where they *can* deliver that unequivocal positive impact in their products. DHH alluded to "really neat ideas" specifically around Basecamp, though he wouldn't reveal spoilers. He had a "peak experience" with an agent recently that convinced him of its power for developers.
HostTell me about that. What was the "peak experience" that really solidified his optimism?
ExpertHe describes a moment where he encountered a really obscure bug in the latest version of Rails, causing a weird artifact when opening the console. He was dreading the deep dive. So, he handed the task to Opus, Claude's latest model.
HostAnd what did Opus do?
ExpertOpus started pulling out a CD bugger – a tool DHH knew existed but wouldn't typically reach for right away. It showed its "thinking" process, hypothesizing, trying things, and even when the first couple of hypotheses didn't pan out, it kept going. Eventually, it pinpointed the exact issue, found the commit where things went wrong, and even provided a patch to work around it until a core fix was available.
HostWait, it debugged a complex framework bug and wrote a patch? That’s wild.
ExpertDHH was blown away. He said he wouldn't have done that himself in that moment, or it would have taken hours. The AI got the problem out of his way. He even compares this level of excitement to playing Yie Ar Kung Fu on a Commodore 64 as a kid, or the first time he published a webpage with Netscape Navigator and realized it was live for the entire world. That "mind-blown" experience of pure technological magic.
HostThat's a powerful comparison, bringing it back to those fundamental moments of technological wonder. But you mentioned he also said it's not *always* like that.
ExpertRight. He admits he's had moments where he gets excited, asks for something seemingly simple, and the AI "makes a mess of things." But the key is that the *ratio* is changing. More often now, it's either solving the problem entirely or providing such a solid draft or blueprint that he can quickly polish it off. He even used a tool called OpenCode to get five different drafts from five different models for a complex task, and all of them produced working solutions.
HostFive working solutions from an open-ended prompt? That's really impressive. It almost sounds like a non-programmer could start building things. Does this mean AI is democratizing access to software creation?
ExpertAbsolutely. This is where Jason Fried, their CEO, comes in. He highlights that while DHH, as an expert programmer, *could* do these things given enough time, the AI is doing something fundamentally different for non-programmers. It’s allowing them to build ideas and produce working applications without understanding "a lick of what's going on underneath."
HostSo, the magic for the expert is speed and offloading toil, but for the non-expert, it's about enabling creation that was previously impossible for them.
ExpertExactly. Jason draws parallels to Excel and FileMaker Pro. These were tools that gave non-programmers the power to create "programs" – monstrous spreadsheets or databases – to solve business tasks, because they couldn't afford a professional programmer or didn't have the skills. AI is now doing that on a much grander scale, democratizing access to creativity.
HostBut there's a caveat, right? These AI-generated programs might not be production-ready or secure.
ExpertA big caveat. Jason is clear that what these models produce today might not have the quality for critical systems like pacemakers or banks. He's seen "horror examples" of AI making things that appear to work but are "leaking like a sieve" or "woefully insecure."
HostSo for businesses, it’s not a blank check to let AI build everything. There’s still a need for human oversight and expertise.
ExpertExactly. However, he argues that for many cases, "it also just doesn't matter." Not everything has that level of criticality. For early-stage programs, or things where the cost of catastrophic error isn't that great, the upside of getting *any* program that didn't exist yesterday is worth the risk. It's the "innovator's dilemma" – everything starts as a "toy" and then grows into robustness.
HostThat makes sense. So, with this rapid evolution, and seeing the internal benefits, how is 37signals approaching AI integration into *their own products* for customers? Are they still waiting, or are they jumping in?
ExpertIt's a nuanced strategy. They've explored it quite a bit in products like Fizzy, their Kanban tool. They tried summary features, for example, for weeks of work. But they found the summaries "adequate but not interesting to read," which pushed them back. They don't want to ship something just because it *can* be done.
HostSo they're not just slapping AI on everything.
ExpertNo. And this is where Jason discusses David's argument, which he fundamentally agrees with: the landscape is changing so fast that users might just bring *their own agents* to the products.
HostWait, so instead of 37signals building AI *into* Basecamp, users will have their own personal AIs that *use* Basecamp?
ExpertPrecisely. Imagine your AI agent signing up for a Basecamp account like a normal user, learning everything about your projects, and then operating within it. They've already seen this internally, with their own agents operating within their Basecamp account. This means 37signals can focus on making their products simpler and clearer, and making it *easier* for external agents to integrate, for example, through command-line interfaces.
HostThat’s a bold take. Most companies are rushing to build their own AI features. 37signals is saying, "We'll open the doors, but you bring your own intelligence."
ExpertIt's a "wait and see" strategy with an open door. Jason is glad they haven't spent thousands of engineering hours building things that might be "undone in a matter of weeks by a better way to do things." He acknowledges there will be a hybrid world: some native AI features within their products, but also users bringing their sophisticated personal agents that connect to everything else they use.
HostThat brings up an interesting point about customer expectations. I've seen the comments – people are really polarized. Some demand AI features, others say, "Don't touch my Basecamp with shitty AI!" How do you balance that?
ExpertIt's a "delicate balance," as Jason puts it. Their approach has always been about straightforward, no-nonsense software. Slathering AI on everything isn't their style. For those who want it everywhere, he believes their personal agents will provide that. For those who don't want to go down that road, or aren't sophisticated enough, 37signals needs to provide "some assistance" and leverage that's "not in your way."
HostIt's the classic "more features versus simplicity" debate, just with AI as the latest frontier.
ExpertExactly. He compares it to adding more tools to a Basecamp project. You can have multiple to-do sections or message boards if you want, but it's still simple for everyone else who just wants the basics. The power is there if you reach for it, but it doesn't clutter the experience for those who don't.
HostThat makes a lot of sense from a product philosophy perspective. But let's broaden this out a bit. We've heard a lot about tech layoffs, with AI often cited as the reason. The idea that AI will replace jobs, especially programmers. What's their take on that?
ExpertJason Fried has a very clear, and somewhat contrarian, view. His longstanding opinion is that most companies tend to be too big in the first place, with too many people on teams. He thinks companies often "look for reasons to let people go that may or may not be true."
HostSo, AI becomes a convenient excuse?
ExpertPotentially. He acknowledges that AI *should* make people more productive, but he also believes people "should have probably been a lot more productive to begin with." He doesn't think you need a team of 12 people working on something relatively small. For 37signals, with about 60 people, AI is seen as a way to develop faster with *fewer* people, but not necessarily to eliminate roles in a company that's already intentionally lean.
HostThat flips the script on the layoff narrative. It's not necessarily AI *taking* jobs, but perhaps companies realizing they were overstaffed and using AI as a justification for efficiency.
ExpertThat's his take. And it extends to things like customer service. He agrees that AI-based customer service can be good for simple, straightforward questions. But he finds it "incredibly frustrating" when you have a nuanced issue and you're stuck in a loop with a bot.
HostOh, I've been there. "No, I'm the bot and I can help you." "No, you're not helping me!" It's infuriating.
ExpertPrecisely. Jason's point is that humans want to talk to humans, especially when they're frustrated. For 37signals, human support is a "massive competitive advantage." They have highly trained, long-term humans on their support team, some for 15 years, who know the product inside and out and genuinely care.
HostSo for 37signals, AI is about augmentation and efficiency, not about replacing the human touch, especially in critical areas like customer relationships.
ExpertAbsolutely. They offer some AI help for quick answers, but you can always email support@basecamp.com and get a human. He says they don't ever want customers to feel like they *must* go through AI and get pissed off to finally reach a human. It's about having both options, but prioritizing the human connection, especially when emotions are involved. People relate to each other, they need to vent sometimes, and that's something AI, however sophisticated, still struggles to truly replicate.
HostSo, if we look back at this whole shift, what are the core takeaways for businesses and individuals trying to navigate this incredibly fast-moving AI landscape?
ExpertI'd say there are a few key insights. First, don't just jump on the hype train. Real value often comes from applying AI to specific "toil" or friction points, not just slapping it onto every feature. Second, the "agent mode" paradigm is a game-changer for productivity, allowing AI to act more autonomously, especially for developers and even non-programmers. Third, consider a "wait and see" approach for product integration, recognizing that users might bring their own sophisticated AI agents, and focus on making your products open to those connections.
HostAnd that leads right into the fourth point, which is about democratizing creation. AI empowers non-experts to build things they couldn't before, even if the output isn't perfect, it's a huge leap.
ExpertAnd finally, don't let the fear-mongering overshadow the incredible potential. While risks are real, the speed of innovation, especially in the last few years, is unprecedented. It's a privilege to be alive at this moment, witnessing what DHH called "an incredible achievement of mankind."
HostAn incredible achievement. And it leaves me wondering: if AI continues to evolve at this pace, will our definitions of "human work" and "human connection" fundamentally change, or will they simply become even *more* valued as the unique domain of people?