73%.

That's the share of all code PRs (Pull Requests) at Spotify that are now directly generated by AI. When this number came out in a recent conversation between the “father of Claude Code” (Anthropic's head of engineering) and Spotify's engineering team, the entire developer community was stunned. Even more staggering are the supporting numbers: 2,900 engineers, 4,500 deployments per day — meaning each engineer participates in over 1.5 production deployments per day on average.

What Does AI-Native Development Look Like?

Many people's understanding of AI coding is still stuck at the “Copilot completes a few lines for me” stage. What Spotify demonstrates is a completely different picture: AI-Native Development.

In Spotify's workflow, engineers no longer write code line by line — they are managers and reviewers of AI systems. A typical development flow goes like this: the engineer describes what feature to build or what bug to fix; the AI Agent understands the requirement, reads the codebase, writes code, writes tests, runs tests, fixes bugs, and finally submits a PR. The engineer's job is to review that PR — confirm the logic is correct, there are no security issues, and it meets team standards — then hit merge.

That's how they achieve 4,500 deployments per day. In traditional development, an engineer writing two to three hundred lines of effective code and submitting one or two PRs per day is considered good; but with AI assistance, an engineer can “manage” multiple AI tasks in parallel, multiplying PR output severalfold. And because AI writes tests alongside code, test coverage is actually higher than when written purely by humans.

The father of Claude Code raised a key change in the conversation: AI coding tools are evolving from ‘editor plugins’ to ‘engineering environments’. Claude Code isn't a completion box inside an IDE — it's a complete Agent environment with direct access to the command line, filesystem, Git, and test frameworks. It doesn't just write code; it runs code, tests code, debugs code, completing the full flow an engineer goes through from receiving a requirement to submitting a PR.

Productivity Up, But Are Engineers Disappearing?

GitLab's survey last week claimed “AI speeds up coding but hasn't improved overall delivery efficiency,” and Spotify's case offers a different answer. The key distinction is: are you treating AI as “smarter autocomplete,” or as an “engineering Agent that can complete tasks independently”?

If you're only using Copilot to complete a few lines, bottlenecks indeed shift to testing, integration, and deployment. But if AI can complete the full closed loop from requirement to PR end-to-end — including writing tests, running tests, fixing bugs — then the entire delivery pipeline accelerates. Spotify's 4,500 daily deployments aren't a myth; they're the natural result of AI permeating the full workflow.

But that doesn't mean engineers are becoming unemployed. Spotify's engineers are actually busier — just busy with different things. They no longer spend large amounts of time writing boilerplate, tweaking APIs, or fixing trivial bugs; instead they spend more time on system design, architecture decisions, code review, and product thinking. An InfoQ analysis from the same period put it well: “Frontend and backend disappear together: AI Coding is rewriting Big Tech engineer roles”. Traditional frontend/backend/testing boundaries are blurring, replaced by end-to-end delivery roles closer to the business.

Of course, this model isn't without risks. The more code AI generates, the greater the pressure on code review — if 73% of PRs are written by AI, the person reviewing must understand the system better than the AI, or they're reviewing blindly. Cloudflare CEO's warning about “a Log4j-level vulnerability every week in the Agent era” isn't alarmist. But the trend is clear: AI coding tool competition has upgraded from ‘who completes code more accurately’ to ‘who can complete engineering tasks end-to-end’. Today it's Spotify; tomorrow more companies will join the AI-native development ranks.