How Spotify's Agents Rewired a 20M-Line Codebase in Two Months

Niklas Gustavsson, who leads software engineering at Spotify and trained as a molecular biologist, once predicted that engineers would stop using IDEs by the end of the year. At the time it sounded implausible on anything shorter than a two-year horizon.

Instead, the shift happened in about two months. Gustavsson found himself no longer using an IDE, describing a change in his workflow unlike anything he'd seen in three decades in the industry -- and the internal experience matched what was visible externally.

Spotify's codebase is split into thousands of components, each owned end-to-end by a specific team responsible for design, implementation, and operation. Historically, teams could review every PR merged into their code, which meant test automation could afford to be a bit loose.

Rolling out AI agents that open PRs automatically meant teams would no longer see every change before it merged. That required building out much stronger test automation so the codebase could survive automated changes safely.

With that verification infrastructure in place, Spotify could point agents at the same safety net. Ideas that once required mobilizing an engineering team for weeks now become working prototypes with real data and real users in an hour or two.

That shift reaches beyond engineering: several senior executives, including one of Spotify's co-CEOs, have built and shipped their own prototypes to the internal app store -- something unimaginable a year ago and now routine.