Background Coding Agents: Context Engineering (Honk, Part 2)

Spotify shares lessons from building a background coding agent ("Honk") that automates large-scale code migrations by generating mergeable pull requests across thousands of repositories.

• •After experimenting with open-source agents (Goose, Aider), Spotify built a custom agentic loop, then migrated to Claude Code for its natural task-oriented prompting and subagent management
• •Claude Code is now their top-performing agent, used for ~50 migrations and the majority of background agent PRs merged to production
• •Key prompting lessons: describe end state not steps, state preconditions, include code examples, define verifiable goals (tests), and do one change at a time
• •Tools are intentionally limited to a "verify" MCP tool, a restricted Git tool, and a Bash allowlist—avoiding code search or docs tools to maximize predictability
• •Larger static prompts are preferred over dynamic MCP-fetched context because they are versionable, testable, and more predictable