April 29, 2026
Codebase of Chaos, meet AI Crew
I built ten custom subagents to tame a 500K-line Clojure codebase
He made 10 AI sidekicks to survive a giant app — and commenters are obsessed
TLDR: A Metabase developer built 10 specialized AI helpers to handle different parts of an enormous app, so the system stops wasting time relearning everything. Commenters are split between calling it brilliant and joking that if software needs a whole robot staff, the real problem might be the software itself.
A developer behind Metabase just dropped a very online confession: their app’s backend had grown into a half-million-line monster, so they built 10 custom AI helpers to keep from getting lost in it. Each helper was trained for one “neighborhood” of the code — like permissions, alerts, or the part that turns user questions into database searches — so the main AI wouldn’t waste time relearning the map every single time. In plain English: instead of one stressed intern trying to memorize the whole company, they hired ten tiny specialists.
And yes, the community reaction is the real show. The strongest vibe is a mix of “this is genius” and “we have officially entered the AI middle-management era.” People are fascinated by the idea of giving each bot a job description, memory, and a specialty, but there’s also side-eye over whether this is clever problem-solving or a sign that modern codebases have become too huge for humans to comfortably hold in their heads. Even the author jumped into the discussion admitting the messiest part is routing overlapping domains — basically, deciding which tiny expert gets the call when two areas blur together.
The jokes practically write themselves: commenters are treating these bots like a sitcom cast, with one person imagining a tiny office full of specialists arguing over who owns the bug. The hot take underneath the humor? If you need ten robot experts to navigate your software, maybe the software is the drama. Still, plenty of readers sound impressed that someone turned AI from a generic chatbot into a team with actual roles — and that idea clearly hit a nerve.
Key Points
- •The article describes a Metabase engineer building ten custom Claude Code subagents for different backend domains in a 500,000-line Clojure codebase.
- •The main problem identified is that a generalist Claude session consumes too much context window capacity repeatedly exploring unfamiliar subsystems.
- •The subagents are organized around domain boundaries such as the query processor, permissions system, database drivers, and notification system.
- •Each subagent is defined as a markdown file with its own system prompt, context window, memory, toolkit, model selection, and persistent storage.
- •An example query-processor agent includes knowledge of Metabase’s middleware pipeline, key source directories, caveats about SQL dialects, and testing or REPL workflows.