How we built ten custom agents to tame our giant codebase

Metabase's extensive backend, consisting of approximately 500,000 lines of Clojure code across various subsystems, prompted the development of custom AI subagents using Claude Code to manage its complexities. These subagents, each with a defined domain such as query processing or permissions, are designed to streamline navigation and reduce the context tax by embedding specialized knowledge into their system prompts. The subagents, created as markdown files, encapsulate domain-specific expertise, notable codebase locations, investigation strategies, and caveats, thereby enhancing efficiency and accuracy in codebase management. For instance, the "mbql-expert" agent understands Metabase's intricate query processing pipeline, enabling it to diagnose issues like incorrect nested query results on Redshift more effectively than a generalist approach. This strategy allows for parallel agent deployment, where multiple specialists can evaluate changes simultaneously without cross-contaminating insights, showcasing a scalable model for managing large, multifaceted codebases.