Spotify engineering blog

Spotify examines how AI-assisted development and faster change exposed weaknesses in content processing, fleet updates, compute capacity, mobile quality signals, and delivery controls. It details incidents, monitoring and prioritization fixes, regional failover changes, and metrics used to assess whether higher velocity is creating software quality debt.

Spotify examines why Bayesian A/B testing is not a single methodology, showing how priors, likelihoods, and stopping rules determine guarantees around peeking, false discoveries, and winner’s curse. It compares Bayesian configurations with frequentist methods and explains why Spotify continues to use frequentist tooling.

Dimitri Mazmanov explains how Spotify’s Portal modes and Claude Code hooks route bulk file reading and boilerplate generation to cheaper worker models. The shunt plugin reduced Claude Code token usage by about 90%, while the post details its routing layers, scripts, benchmarks, latency trade-offs, and limitations.

TL;DR: LLM predictions can stand in for human outcomes in A/B tests, but only by assumption, not by design.

Companies like Spotify need vast quantities of data accessible at low latency for online services and, ...

Over the past two months, podcast creators have experienced a series of reliability issues on Spotify. The post appears to discuss incidents affecting content ingestion and podcast video reliability.

Spotify's goal for AI is bold, but simple: make it as intuitive and indispensable to daily work as email or...

At Spotify, data problems used to follow a specific pattern: people would search for the right dashboard, data source, or domain knowledge before they could make progress. This post introduces a context layer for Spotify's data assistant that encodes domain expertise to improve how users find and use the right information.

At Code with Claude, Spotify’s chief architect shared how we make both teams and AI agents more effective.

TL;DR LLM evals, automated judges that assess relevance, coherence, and quality at scale, are a powerful new...

Turning OpenAPI spec and Markdown files into a conversational ads management tool — no compiled code required.

How we used Honk, Backstage, and Fleet Management to ease the pain of migrating thousands of datasets.

This post discusses how AI agents are transforming the way we build software and how developers think about their work. It summarizes Spotify’s live session with Anthropic on agentic development, covering practical examples and the implications for developer workflows.

What if we could identify interesting listening moments from your year, and tell you a story about them? This post describes the engineering work behind generating 2025 Wrapped highlights, explaining how Spotify identifies moments and assembles them into a narrative.

Spotify describes the multi-agent architecture they built to improve advertising systems, emphasizing that the goal was to fix structural problems rather than to ship an "AI feature." The piece explains how multiple cooperating agents solve complex ad tasks more reliably and integrate with existing infrastructure.

Part 2 examines the tooling that powers Spotify's release process, giving a behind-the-scenes look at the systems that make releases possible. It covers the automation, pipelines, and orchestration used to reliably build, test, and ship the Spotify app.

TL;DR Established in 2022 as a way to help support the great open source ecosystem projects that Spotify...

This post lays out the technical and practical reasons for keeping personalization and experimentation on distinct tech stacks. It argues that separation reduces coupling, lets each domain optimize for its own requirements (such as ML serving vs. rapid experiment iterations), and improves reliability and developer velocity.

The article describes the system Spotify built to ensure AI coding agents produce predictable and trustworthy code through robust feedback loops. It focuses on the design, monitoring, and iterative feedback mechanisms used to validate and improve agent outputs.