Cursor now ships a coordinator agent that runs your codebase for months while you're not looking
Cursor Projects, launched in beta on 10 September 2026, adds a cloud-based coordinator agent that plans work, delegates it across thousands of parallel subagents, and keeps running on recurring tasks for months without a developer prompting it — a step change from AI pair-programming toward AI project management, billed at API rates on top of a paid Cursor plan.
20 September 2026
Cursor launched Projects in beta on 10 September 2026, and it’s a different kind of feature to the autocomplete-and-chat tools most people still picture when they hear “AI coding assistant.” Projects adds a coordinator agent that doesn’t write code itself — it plans the work, delegates pieces of it to cloud subagents (Cursor describes dispatching thousands of them across a single project), and hands finished pull requests back for a human to check. Crucially, it maintains context over months, runs on a cloud machine rather than a developer’s laptop, and can perform recurring work without being re-prompted each time.
There’s no separate Projects fee — it runs on Cursor’s existing Cloud Agents infrastructure, billed at API pricing for whichever model you choose, on top of a paid Cursor plan.
The shift this represents
Most AI coding tools to date have automated a task: write this function, fix this bug, refactor this file. A coordinator agent automates something closer to project management — deciding what needs doing, sequencing it, delegating it, and persisting the plan across weeks or months rather than a single session. That’s a meaningfully different capability, and it’s the direction several major AI coding vendors are converging on: less “AI helps you code faster,” more “AI runs an ongoing work queue that a human periodically reviews.”
What it doesn’t change
The coordinator still hands results back “for you to check” — Cursor’s own framing keeps a human in the review loop by design, which is worth noting given how much scrutiny AI-generated code has come under this year for security gaps and unreviewed edge cases. Persistent, delegated, multi-agent work at this scale doesn’t remove the need for someone accountable for what ships; if anything it raises the stakes on having a review process that can actually keep pace with work being generated continuously in the background rather than in reviewable, human-paced chunks.
So what
If you’re commissioning a build and evaluating whether “AI-assisted” agencies are meaningfully different from each other, this is a good question to ask directly: is the team using tools like this as a force multiplier inside a process with real review gates, or as a way to generate more code faster with less oversight? The tools now exist to run genuinely autonomous, long-horizon coding work — which makes the quality of the process wrapped around them the actual differentiator, not access to the tools themselves. That’s the standard our AI-assisted development work is built to hold.