Cursor 2.0 ships Composer, multi-agent UI

Cursor is selling the workflow, not just the model: faster turns, parallel agents, and a tighter test/review loop matter more to most teams than whatever base weights sit underneath. If the Kimi theory proves out, it just underlines how quickly AI IDE competition is shifting from model invention to productization and control of the harness.

• –Cursor’s own post frames Composer as a frontier MoE model trained with RL on real codebase tasks and semantic search, so the official story is still “we built the model for this job.”
• –The reverse-engineering claim matters because it points to a broader pattern: vendors can wrap strong open or third-party models, then differentiate on editor integration, tool use, and latency.
• –The multi-agent UI and isolated worktrees are the most defensible part of the release; they address real coordination problems that single-chat copilots don’t solve well.
• –If Cursor is routing to Kimi-derived weights, licensing and transparency become product issues, not just forum drama, especially for customers paying for trust and stability.
• –For developers, the useful takeaway is that the winning coding assistant is increasingly a system: model + harness + browser + review loop, not just raw benchmark scores.