The coordination tax: what happens when your AI agents can't see each other

You start with one agent. It's great. You add a second. Still fine. By the third, you notice something: they're duplicating work. By the fifth, they're actively fighting — overwriting each other's files, claiming the same task, producing conflicting outputs. This is the coordination tax. And it scales faster than your agent count. The problem isn't the agents Modern coding agents — Claude Code, Codex, Cursor — are individually excellent. The problem is they operate in isolation. Each one thinks it's the only agent in the system. There's no shared state. No shared task board. No presence awareness. No way for Agent A to know that Agent B is already working on the auth module. What coordination actually looks like We run 9 agents on a single Mac Mini building a product together. Here's what we needed: Task ownership : An HTTP endpoint where agents claim tasks. If it's claimed, you can't take it. Simple mutex. curl -X POST localhost:4445/tasks \ -H "Content-Type: application/json" \ -d '{