Cursor vs evaluate

Cursor's last two months are almost entirely about cloud agents rather than the editor. Agents now launch from iOS, iPad, and Slack, run in isolated VMs, and boot from environments Cursor rebuilds in the background. Auto mode became Cursor Router, which dispatches each request to a frontier or price-efficient model and bills at the routed model's rate. Team plumbing matured alongside it: marketplaces, Team MCPs, plugins, and automations.

The direction is making the agent rather than the editor the unit of work, then removing every source of latency and friction around it. Environment setup went from a manual step to a reusable snapshot to a continuously rebuilt warm pool; model selection went from a picker to a router with admin controls; surfaces went from desktop to phone, tablet, and chat. Nearly every release pairs the user-facing feature with an admin or team control, which points at enterprise buyers as the audience being served.

The next moves most likely target cloud agent throughput and spend controls — finer routing or more environment reuse — rather than another client surface, since mobile and Slack have each had two releases already and the newest work concentrates on what happens before an agent's first token.