Metabase engineer Bryan Maass ran into a familiar problem. Their 500,000-line Clojure backend kept eating Claude's context window. Each subsystem, from the query processor to the permissions system, has its own idioms and hidden knowledge. Every time Claude needed to understand a new domain, it would explore, grep, and read files. Tokens burned on repeated discovery. His fix: ten specialized subagents using Claude Code, each one a domain expert for a different corner of the codebase. The approach is straightforward. Each subagent is just a markdown file, about 150 lines, packed with domain-specific context that Claude can't easily infer from code alone. The mbql-expert agent knows the query processor's 68-stage middleware pipeline, the four phases it runs through, and critical caveats like "BigQuery is not standard SQL." The permissions-expert understands multi-granularity graph handling, row-level security, connection impersonation. Each agent gets its own context window, system prompt, memory directory, even its own model selection. Maass used Claude itself to help write these "job descriptions", describing what an expert in each domain would need to know. No more "context tax." Instead of one generalist session burning tokens to relearn a subsystem every time it switches domains, each subagent starts with the knowledge it needs. The files follow a consistent structure: domain knowledge, codebase locations, investigation patterns, caveats, testing strategies. Maass made all ten markdown files publicly available. There's no formal Model Context Protocol here. No agent-to-agent communication. He achieved multi-agent coordination by defining job descriptions that give each agent isolated context for its subsystem. Each one just knows its domain cold.