KorvaJoin the waitlist
Tech

When Building Gets Cheap, Taste Gets Expensive

With agents writing thousands of updates, a software team says saying no matters more than shipping fast.

AI image
Y Combinator · The Rise Of AI Coworkers

Lauren Tan, a software engineer building AI coworkers at a startup called Grockbot, and her colleague Roman Ugarte were talking through what happens when a team can spin up 100 coding agents at once.

The interviewer offered a prediction about that abundant future: "Maybe the scarce resource is going to be more that higher executive function of what to do and why."

The interviewer suggested the gap would probably not be an execution failure but a failure to specify intent, goals and measures of success. The interviewer added that companies would need clearer targets and owners for that work, especially in areas like marketing where improvement is harder to define than faster software.

Ugarte said the team is already feeling that shift. Tan puts up about 4,000 proposed software updates a month, Ugarte said, and the bottleneck is less engineering time than agreement on the roadmap, the vision, and what good or great looks like.

Proposals now arrive before decisions do. Ugarte described meetings about a future change when someone says he already has a proposal built from weekend work, and warned that merging everything would create a "Frankensteiny product" without a consistent vision.

So the team splits routine repair from choices that need judgment. Tan set up a system called Sandy that takes in thousands of customer comments a day and quietly fixes small bugs, sometimes merging and shipping at 3 a.m. without discussion.

Changes to buttons and screens people see every day get the opposite treatment. The team tries them internally and sits with a change longer than feels comfortable, and drops anything without clear internal support, even when reactions are mixed.

Ugarte put it this way: "even though you can go really fast, it doesn't mean that you should just ship every feature you can think of." Ugarte added that human judgment and taste come in there, and he does not think that will go out of style. As long as people are building software for humans, Ugarte argued, people should be thinking about an intentionally designed experience because humans know each other best.

Questions people ask

What does 4,000 PRs a month actually mean?

PR stands for pull request, a proposed change to a software project that others can review before it is added. Four thousand a month means Tan's agents are proposing more than 100 changes a day, from tiny fixes to possible features.

What is dogfooding in practice?

Dogfooding means using your own product the way customers do, to feel annoyances before they do. For Grockbot, that means living with a moved button or new screen for a while, and killing the change if coworkers find it confusing or irritating.

Why let AI fix bugs automatically but not design features?

Small bugs have a clear right answer, so agents can fix them with little risk. Feature choices involve tradeoffs, habits, and taste, and a technically correct change can still make the product messier, harder to learn, or less coherent.

More in Tech