Spot on that it's tool engineering over model capability. The speculative edits trick is a perfect example - it's not about having a smarter model, it's about being clever with what's already there. Your multi-tool split makes sense too, Cursor's architecture is optimized for inline speed while Claude Code's strength is longer autonomous tasks. The 90-minute RL loop is wild though - it's not really learning your codebase, it's learning your accept/reject patterns across all users. Still impressive at 400M requests/day. How stable has your tool split been as each one keeps shipping new capabilities?