Cursor launches Projects, a coordinator agent that delegates work to cloud subagents

Cursor launched Projects in its changelog, a new mode where a coordinator agent plans a piece of work, delegates it to subagents that write the code, and hands the finished output back for review. Cursor says a Project runs on its own computer in the cloud so closing the laptop does not stop it, and the coordinator can spawn a local agent when a step needs to run on the developer's machine.

What the coordinator does

Cursor describes the coordinator as an agent that does not write code itself. It plans the work, delegates it to agents that implement each piece, and brings the finished work back for the developer to check. Cursor says the coordinator creates and manages agents on the user's behalf and can run as many in parallel as the work needs, with the changelog naming "thousands of subagents" as the target scale.

Each Project keeps a set of files that sync across every cloud and local machine its agents use. The changelog says agents add research, artefacts, and what they learn about the codebase and about the developer's preferences. If one agent works out how to test a service, later agents in the same Project can read those notes and skip the discovery step.

Triggers instead of prompts

The second half of the launch is scheduled and event-driven work. Cursor says the coordinator can watch a Slack channel, run on a schedule, or follow all pull requests, and start work when a signal comes in. The example the changelog gives is connecting Slack, pointing the coordinator at a bug-report channel, and letting it delegate each incoming bug to a subagent without a human prompt.

That is the piece that separates Projects from Cursor's earlier agent runs. An ordinary Cursor session ends when the developer closes the tab. A Project keeps running against its context and its triggers, so a bug reported in Slack over the weekend can have a draft fix waiting on Monday.

What the changelog does not say

Cursor's post does not give pricing or a plan tier for Projects. It does not name the models the coordinator runs on, or the models the subagents use, and it does not put a number on the concurrency limit inside a single Project. The scale claim is a soft "thousands of subagents" rather than a per-Project cap. Teams that pay Cursor per model call will only learn the cost of running one of these against their own bill.

For a team that has been juggling one long-running Cursor agent per feature, the change is that context, files, and preferences now belong to the Project rather than to the session, so a second engineer joining next week does not restart the onboarding. For a team that has been running scheduled coding tasks through GitHub Actions or a homemade Slack bot, the coordinator is Cursor's answer to that pattern, wired into the same editor and the same set of subagents. Whether the delegated model calls cost less than the equivalent hand-run Cursor session is the question the pricing page does not answer yet.

Source

Projects on the Cursor changelog.