Key Takeaways

  • Ryan Carson runs 10 to 15 autonomous Devin agents in the cloud for Untangle, shipping dozens of pull requests every day.
  • Raw code volume does not create market demand. Untangle stalled when Carson built it as a consumer AI divorce app because people facing divorce want human lawyers, not chatbots.
  • Carson saved Untangle by leaving his desk to pitch family law attorney Renee Bower, which triggered an immediate pivot to B2B software for law firms.
  • Claire Vo intentionally restricts feature shipping velocity because pushing multiples of code output does not yield multiples of product quality.
  • Frontier models cannot evaluate market need or exercise product taste, making direct customer interviews irreplaceable.

The Trap of Token Maxing

It is easier than ever to drown in your own output. When you can run a fleet of 10 to 15 autonomous Devin agents in the cloud to write code, test features, and open dozens of pull requests while you sleep, your output explodes. But shipping code fast does not mean you are building something people want.

Founders fall into token maxing because writing software feels like progress. Watching agents generate thousands of lines of code provides the illusion of building a valuable business. Carson experienced this trap firsthand with Untangle: “I did what all founders do, right? You know, launched a version of Untangle, which is the product, thought it was for consumers, went out, nobody. Turns out, nobody wants to use AI for divorce. They want lawyers, people.”

Generating 50 pull requests a day for a product nobody wants just gets you to the wrong destination faster.

How One Meeting Saved Untangle

Carson did not fix Untangle by prompting his agents to rewrite the user interface or add new consumer features. He fixed it by closing his IDE and booking a real conversation with someone in the industry.

He scheduled a meeting with family law attorney Renee Bower. Carson walked in and laid out the product with complete candor: “I was like, 'Okay, I'm going to get a meeting with a lawyer and just see what they think of this thing.' And I met with this amazing lawyer named Renee Bower and I was like, 'Do you love this? Do you hate it? Is this like an enemy? Is it a friend?' And she said, 'I love it.'”

That single conversation revealed that while consumers going through a divorce wanted human representation, divorce attorneys needed software to streamline their case operations. The meeting triggered an immediate pivot to B2B legal software. No prompt or automated agent workflow could have surfaced that insight from behind a screen.

Why Frontier Models Lack Taste

Automated coding tools are execution engines, not strategists. They can write functions, but they cannot feel market friction. Vo actively restrains how much product she pushes to production for this exact reason: “I really don't ship more than I think the market wants. I really try to constrain my output, not on quality, not on bugs, but because I don't think I get multiples of quality off of multiples of output.”

Carson agrees that founders expecting AI to direct product strategy are mistaken: “I actually think we're nowhere near any frontier model having the intelligence to know what to ship. It's like a million miles.” When software creation becomes free, the scarce resource is knowing what to build.

As Carson puts it: “People are not getting out of their chair enough and actually talking to real people... because we can ship so much and build so much digital product, we're not getting out of our chairs enough.”

What to Do With This

Audit your calendar for the coming week. If you have spent 40 hours prompting coding agents and zero hours talking to buyers, pause your agent workflows for two days. Book three 20-minute user discovery calls with practitioners in your target industry before you merge another pull request.