1,500 PRs Later: Spotify's Journey with Our Background Coding Agent (Part 1) (engineering.atspotify.com)

Spotify extended its Fleet Management system by integrating background AI coding agents that have already generated and merged more than 1,500 pull requests. Instead of writing brittle, hand-coded source-to-source transformation scripts (their Maven updater grew to ~20k lines), engineers can describe migrations in natural language and let an agent produce PRs that are reviewed and merged by existing Fleet workflows. The approach has cut manual effort dramatically — reported time savings of 60–90% for migrations — and is now used for nontrivial work like modernizing Java to records, upgrading Scio pipelines, migrating UI components in Backstage, and schema-safe config edits across thousands of repos. Since mid‑2024 roughly half of Spotify’s PRs were automated by Fleet-style tooling, and AI has broadened the class of changes that can be automated. Technically, Spotify wrapped LLM-driven agents in an internal CLI and Model Context Protocol (MCP) layer so prompts, local formatting/linting, LLM diff evaluation, GCP logging, and MLflow traces are all integrated while remaining pluggable across models and agents. The agent can be invoked from Slack, GitHub, or IDEs via an interactive front end that builds prompts and hands off tasks to the background coder. Key challenges remain — latency, output unpredictability, validation, sandboxing, and compute cost — and Spotify is prioritizing context engineering, feedback loops, and robust guardrails to make agent-driven, large-scale code maintenance reliable and safe.