Spotify says 73% of PRs are AI-assisted now

Spotify says it’s shipping 4,500 production deploys a day and that 73% of pull requests are now AI-assisted. VP of Engineering Niklas Gustavsson also details running multiple Claude agent sessions in a 20M+ line monorepo—and why verification is key.

Spotify says 73% of PRs are AI-assisted now

TL;DR

  • Scale at Spotify: 4,500 production deploys/day; 73% of PRs AI-assisted
  • Workflow: 5–10 Claude sessions in tmux; one per git worktree; agents run in background
  • Codebase context: Agents operate inside a 20M+ line monorepo; reported to “work well”
  • Migration codemods: Thousands of edge cases; large API surface makes static rewrites unreliable
  • Judge-based evaluation: PR success rate increased from ~25% to 80% after adding a judge
  • Verification focus: Rebuilt test automation around verification to enable supervision over repetitive manual work

In a ClaudeDevs video, Boris sat down with Spotify VP of Engineering Niklas Gustavsson for a discussion centered on AI-assisted software work at Spotify. Gustavsson states that Spotify ships "4,500 production deploys a day" and that "73% of PRs are now AI-assisted."

He also mentions keeping "5 to 10 Claude sessions" running in tmux, one per git worktree, with agents working in the background inside a "20M+ line monorepo." Gustavsson adds that he expected agents to struggle at that scale, but that it has "worked well."

Another part of the interview focuses on Spotify’s migration codemods, which Gustavsson says grew into "thousands of lines of edge cases." He claims the codebase has too much API surface for static rewrites to handle cleanly, and that early LLMs "barely did better." After adding a judge, he says PR success rose from "about 25%" to "80%."

Gustavsson also argues that verification is "the single most important thing" when agents are used, describing it as an area where many companies underinvest. According to his account, Spotify rebuilt test automation around verification so engineers can supervise agents instead of manually carrying out repetitive tasks.

Source: ClaudeDevs

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community