Cursor's Projects beta puts a coordinator agent in charge of your other agents

Cursor shipped Projects in beta on September 10, 2026, and the headline idea is a change of job description: the agent you talk to stops writing code.

Instead, a Project gives you a coordinator agent. Per Cursor's changelog, a Project "maintains context over months of work, delegates tasks to thousands of subagents, and performs recurring work without being prompted." The coordinator plans the work, hands implementation to other agents, and brings the finished result back for you to check.

What's actually new

Three things separate this from "open four agent tabs and hope."

It runs in the cloud, not on your laptop. A Project gets its own cloud machine, so closing the lid doesn't stop the work. TestingCatalog reports that local agents can still be launched when a task needs your hardware to test against.

Context is treated as an artifact. Shared files sync between the Project's cloud machine and your local ones, carrying research, codebase knowledge, testing instructions and developer preferences into later tasks. The onboarding cost of explaining your repo gets paid once rather than once per session.

It can start work without you. Subscriptions let a coordinator watch a Slack channel, run on a schedule, or follow your pull requests — reacting to a bug report or a failing CI run rather than waiting for a prompt.

Cursor names three patterns it sees internally: building a feature, running an incremental migration, and ongoing maintenance work it calls "gardening."

The numbers, and why to hold them loosely

Cursor's internal figures are striking: new Projects users merge 30% more PRs, heavy users merge roughly six times as many, and one internal design-system Project is expected to touch 20 to 100 PRs a day.

Treat those as vendor claims. As AI Weekly notes, no baseline, cohort size, or time window was published alongside them. "Merged PRs" is also a throughput metric, not a quality one — an agent fleet that opens more PRs than you can meaningfully review has moved your bottleneck rather than removed it.

The same write-up flags two real limits: Projects reportedly underperforms on quick one-shot tasks, and multi-week runs carry context-drift risk that demands serious CI discipline. Notably, the coordinator is blocked from pushing to main branches on its own.

What it means if you're learning to build with AI

The skill being priced up here is not prompting. It's decomposition and review — cutting work into pieces that don't collide, and judging output you didn't write.

If that sounds like engineering management, that's the point. You don't need a Projects subscription to start practising it: give one agent one narrowly scoped task, insist on tests, and review the diff before it merges. Those habits are what make a coordinator useful instead of a firehose.