The per-agent code wrapper approach breaks down when managing dozens of agents each with their own tools, MCP servers, and credentials; policy needs to be centralized into a single layer through which all calls pass.