Claude Code just shipped a massive upgrade to its agent system: Agent Teams. This isn’t a small iteration on the old task + sub-agent model. It’s a fundamentally different execution model that allows 3–5 independent Claude Code instances to collaborate on the same project, share context, exchange messages, and coordinate through a shared task system. I spent time digging into the logs, tracing model calls, and inspecting the filesystem changes behind the scenes. After a lot of back-and-forth…