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

This post details the evolution of Spotify's background coding agent ('Honk') from early experimentation with open-source tools and a homegrown agentic loop to leveraging Claude Code. It introduces the concept of 'context engineering' for coding agents, emphasizing the importance of well-crafted prompts and the right tools for reliable, large-scale code migrations. Key contributions include: 1. Identifying challenges with early agents (scalability, mergeability, context window limitations). 2. Describing the homegrown agentic loop and its limitations. 3. Highlighting the benefits of Claude Code for natural language prompting and dynamic task management. 4. Providing hard-won lessons in prompt engineering (tailoring to agent, stating preconditions, using examples, defining end states, single changes, agent feedback). 5. Detailing the tools provided to the agent (verify, Git, Bash) and those intentionally excluded (code search, documentation).