Spotify's journey with background coding agents focuses on automating code maintenance through context engineering. After initial success with open-source tools like Goose and Aider, the team faced challenges scaling these agents for larger migrations across thousands of repositories. Writing effective prompts became a significant hurdle, especially when the goal was to produce reliable, mergeable pull requests. The homegrown "agentic loop," built on large language model (LLM) APIs, allowed for iterative editing but struggled with complex, multi-file changes and user input limitations. To improve performance, the team adopted Claude Code, which can interpret high-level goals and manage tasks more dynamically. This agent has been used for around 50 migrations and has successfully merged the majority of background agent pull requests into production. The experience highlighted the difficulties of prompt engineering. Common pitfalls included overly generic prompts that left too much to interpretation and overly specific ones that faltered when faced with unexpected scenarios. Successful teams tailored their prompts for Claude Code, stated preconditions to avoid impossible tasks, and provided clear examples to guide the agent's output. Spotify emphasizes the importance of writing prompts that define the desired end state and encourage iterative improvements. They favor larger, static prompts for better predictability and version control, while also limiting the agent's tools to reduce complexity and potential errors. Currently, the agent has access to a verification tool that runs formatters, linters, and tests, making it easier to integrate with various build systems across their extensive repository landscape. This focused approach helps streamline the coding process and enhances the agent's reliability in producing effective code changes.