Hey Stas, thanks for the detailed feature request.
Some of what you described already works. Cursor has custom subagents. You create files in .cursor/agents/*.md, give each one its own role and instructions, and pick a model for each one, like a cheaper model for a reviewer and a stronger model for an architect. Then the main agent can delegate tasks to them. Your team setup idea frontend, backend, security, reviewer can be done this way. Guide: Subagents | Cursor Docs
What’s not available out of the box yet is a truly long-lived team of agents that keeps working together across many iterations under one conductor and can message each other. The team is actively exploring this, so keep an eye on release updates.
Related threads where you can add your vote: