Build1 publisher2 min readPublished
One engineer reviews the 20 to 100 daily pull requests Cursor's design-system Project touches
Cursor's Projects beta gives each body of work a coordinator agent on its own cloud machine. Its internal deployment shows where the review load lands, and its productivity figures arrive without a baseline.
The Engineer · Build desk
What happened
- Cursor launched Projects on September 10th, a beta persistent coordinator agent that directs coding work across cloud and local machines, announced in a thread on X.
- Cursor's launch post says the coordinator does not ordinarily write code: it plans the work, creates subagents for research, implementation and testing, and collects their output for review.
- Each Project runs on a dedicated cloud computer so work continues after the laptop closes, and the coordinator can start an agent on the developer's own machine to test against a local setup.
- Work can start without anyone typing, through what Cursor calls subscriptions: watching a Slack channel for bug reports, following pull requests, reacting to CI failures, or running on a schedule.
- Cursor's internal design-system Project is expected to touch 20 to 100 pull requests a day, with an engineer reviewing its work.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Review stays with the human. A lead adopting Projects is deciding who sits in front of the merge queue for a fleet's output, and how many bodies of work one reviewer can hold at once.
- cost With no published account limits and no compute price for delegating to thousands of subagents, a buyer finds out what fan-out costs after the coordinator has already run it.
- exposure A subscription puts the trigger for agent work in the hands of anyone who can post in the watched channel or open a pull request, so access control on those inputs now governs what agents start doing.
- constraint Durable project files only compound where an agent can actually run the tests, and in a repository whose setup is not reproducible each new agent rediscovers what the last one learned.
Across an eight-hour day, 100 pull request reviews works out to one every 4.8 minutes, and 20 works out to one every 24 [19]. Cursor says its engineers have used Projects for several months, including migrations spanning hundreds of pull requests, design-system maintenance and the development of Projects itself [8]. The design-system Project extracts reusable components from new pull requests and creates a lint rule after it encounters the same mistake twice [9]. That is the right instinct: a review comment that recurs becomes a check that runs without a person in it.
Project files synchronize across the cloud machine and the local one, carrying plans, demos, research, testing instructions and lessons about the codebase from one agent to the next [6]. Cursor says one coordinator can delegate to thousands of subagents, and the launch materials do not specify account limits or the compute costs attached to that scale [4].
Cursor says new Projects users merge 30% more pull requests, and that users who primarily work through Projects merge six times as many [11]. Six times is a 500 percent increase, roughly 17 times the gap claimed for new users [20]. Neither number comes with a sample size, a measurement window or a baseline [12]. The heavier-use group is defined by its adoption of Projects, so the 6x describes a cohort. It does not identify a cause [13].
For the 30% to transfer to another team, three things would have to hold: a measured pre-adoption merge rate in the same repository, a review policy that did not change across the window, and merged pull requests that track shipped work even when the work is sliced into pull requests differently.
The one randomized figure in the record runs the other way. In a 2025 METR study, 16 experienced open-source developers completed 246 tasks with AI tools allowed or prohibited, and allowing the early-2025 tools increased completion time by an estimated 19% [14]. Cursor was among the tools used, the sample was small, the study predates Projects, and it did not test persistent fleets working asynchronously [15].
Michael Truell set out the direction on February 26th, writing that AI coding was progressing from autocomplete, through synchronous agents, toward fleets that operate for longer periods with less direction [16]. When Cursor raised its 2023 seed round, Truell described the goal as making programming "an order of magnitude faster" [17].
What to watch
- Whether Cursor publishes a baseline merge rate, sample size and measurement window behind the 30% and 6x figures.
- Whether the beta gains per-account subagent limits or published compute pricing for coordinator fan-out.
- Whether any randomized trial measures asynchronous agent fleets rather than single-session assistants.