Spotify Honk: 1,500+ merged agent PRs on Fleet Management
Part 1 of Spotify’s background coding agent: prompt-defined fleet shifts, pluggable agent CLI, 60–90% migration time savings, and new validation needs.
Source: 1,500+ PRs Later: Spotify’s Journey with Our Background Coding Agent (Honk, Part 1)
Fleet Management already automated simple transformations at huge scale; complex migrations drowned in AST/regex expertise (one Maven updater >20k LOC). Honk replaces the transformation script with a prompt-driven agent while keeping targeting, PR, review, and merge infrastructure.
Internal CLI wraps agents with formatting, lint MCP, LLM-as-judge on diffs, tracing—swappable models without exposing guts. 1,500+ merged production PRs on real migrations (Java records, Scio upgrades, Backstage UI, schema-sensitive config). Also exposed via MCP from Slack/GitHub for ad hoc tasks.
Stated open problems: latency, unpredictability → need new validation; safety/sandboxing; cost at scale.
Why it matters for code review
Shows the winning enterprise pattern: agents inside an existing PR machine with verification hooks, not a parallel shadow process. Review burden becomes the next bottleneck once generation works.