Tech • IA • Crypto
How Spotify runs agents across 20M+ lines of code, with Niklas Gustavsson
9.4/10TL;DR
AI agents and large language models are rapidly transforming software development at Spotify, enabling massive gains in productivity, automation, and experimentation.
KEY POINTS
Shift away from traditional coding tools
Developers report abandoning conventional IDE-centric workflows within months as AI agents matured. Tasks that once required manual editing are now handled end-to-end by models, eliminating the “last mile” of code fixes and fundamentally changing daily engineering practices.
Breakthrough driven by advanced models
Early large language model experiments struggled with reliability, but newer generations marked a turning point. Systems evolved from acting as “smart autocomplete” to fully capable problem solvers, handling complex coding tasks with minimal prompt engineering and significantly higher success rates.
Massive codebase automation at scale
Spotify faced a codebase growing seven times faster than its engineering staff, creating a maintenance bottleneck. To address this, it built internal infrastructure to automate large-scale code changes across thousands of repositories, replacing slow, manual migrations performed by hundreds of teams.
From scripts to AI-driven agents
Traditional automation relied on deterministic scripts that became unwieldy due to edge cases in large codebases. AI agents replaced these approaches, enabling more flexible and scalable transformations. Internal systems evolved through multiple iterations into a platform capable of orchestrating automated changes across millions of pull requests.
“Honk” platform and agent architecture
Spotify’s internal system, Honk, runs AI agents in Kubernetes environments with access to development tools and continuous integration pipelines. Agents can execute code, run builds on Linux and macOS, and validate outputs autonomously, enabling closed-loop development without human intervention.
Verification and testing become critical
As automation increased, investment in testing infrastructure became essential. Teams strengthened automated testing because code changes are now often merged without human review. Reliable verification pipelines ensure quality despite high deployment velocity.
High deployment velocity with maintained quality
The company now performs roughly 4,500 production deployments per day while keeping quality metrics stable. Faster iteration cycles allow ideas to move from concept to production in as little as an hour, significantly accelerating product experimentation and feedback loops.
Measurable productivity gains
Internal metrics show a 75%+ increase in pull request frequency, with about 73% of PRs generated by AI tools. These gains are substantially larger than historical improvements in developer productivity, which were often only a few percentage points.
Linking engineering output to business impact
Efforts are underway to connect AI-driven development metrics to user value and revenue. This includes mapping code changes to product features, experiments, and A/B test outcomes, allowing clearer attribution of business impact from engineering work.
Standardization boosts AI effectiveness
Consistent codebases, shared frameworks, and standardized tooling significantly improve agent performance. Uniform patterns make it easier for models to interpret and reuse existing code, while fragmented systems reduce effectiveness.
Rise of AI-powered prototyping across roles
AI tools have enabled not only engineers but also designers, product managers, and executives to build working prototypes. Internal platforms allow users to create and share fully functional features in hours, dramatically reducing the barrier to experimentation.
Changing role of engineers
Engineers increasingly focus less on implementation and more on problem-solving, system design, and exploration. AI handles execution, while humans guide direction, evaluate outcomes, and identify new opportunities.
CONCLUSION
AI-driven development is shifting software engineering from manual implementation toward automated execution and rapid experimentation, with organizations that invest in testing, infrastructure, and standardization gaining the most advantage.