Cursor rolled out Projects in beta on September 10, open to all users. It targets work that spans multiple pull requests and runs for weeks or months at a time: building a new feature, migrating a framework, or standing up an entire app from scratch.
The setup has two layers. On top sits a coordinator agent that writes no code itself — it only hands out work to the agents below it. Because the coordinator only delegates and never executes, Cursor says it never gets stuck on a single task and can take on new instructions at any time. Below it are the agents that actually write code; the company says one project can direct thousands of them.
"It doesn't write code itself but directs other agents that do."
Three Capabilities Behind It
Projects rests on three capabilities:
- Cloud-first: projects run on dedicated servers, so closing your laptop doesn't interrupt them;
- Shared context: files sync between the cloud and local machines, so later agents can reuse what earlier ones already learned;
- Subscriptions: a project can watch a Slack channel, run on a schedule, or react automatically to a PR's status.
Cursor lays out three typical uses. For new features, multiple agents research the existing system in parallel before splitting off to build different components. For migrations, teams use it to swap out frameworks and styling systems across hundreds of PRs, one step at a time. For maintenance, a coordinator keeps a continuous watch on code quality and regressions.
The Design-System Example
The most concrete example in Cursor's blog post is design-system maintenance: a coordinator scans every new PR, pulls out components that belong in the design system, and adds a lint rule the second time it sees the same mistake. By Cursor's estimate, a project like this handles 20 to 100 PRs a day.
Translate that volume into human terms. Assume a careful review of a mid-sized PR takes 15 to 30 minutes: at 20 PRs a day that's 5 to 10 hours, which one person can just about handle; at 100 a day it's 25 to 50 hours, which at eight hours a person requires three to six reviewers working full-time. Once machines can produce this fast, the bottleneck shifts to review — unless a team accepts that some PRs merge after nothing but automated checks. Cursor doesn't say what share of the PRs Projects produces get a human review before merging.
Where the Six Times Comes From
Cursor gives two performance numbers: internally, new users merge 30% more PRs; users who rely mainly on Projects merge six times as many. Both figures come from Cursor's own employees, and both measure merged PR count only — they say nothing about PR size, rework, or rollbacks, and the blog post doesn't specify what baseline the 30% is measured against. No data from outside teams is available yet.
On pricing, the blog post doesn't say whether Projects is billed separately, or how running thousands of sub-agents is metered. Keeping hundreds or thousands of agents running in the cloud for long stretches won't be cheap, and that's a detail Cursor still needs to spell out.
Two Steps Now Connected
Seen against Cursor's moves over the past six months, Projects is the next link in a chain. When it shipped version 3.0 in April, Cursor repositioned its IDE as an interface for managing agents; in early September, it let cloud agents run inside a company's own machines. Projects connects the two: a person manages the coordinator, the coordinator manages the agents doing the work, and those agents can run in the cloud or on a company's own servers.
OpenAI's Agents API, launched the same day, also lets a primary agent split tasks among sub-agents, but the two companies are selling different things: OpenAI sells developers a foundation for building their own agents, while Cursor sells a finished product that does the team's work directly. Projects sits in the left-hand sidebar, and Cursor recommends using it only for work that spans multiple PRs or runs over an extended stretch of time.
Sources: Cursor's official blog, CocoLoop, MarkTechPost; Cursor's blog post was used to verify the sub-agent scale, the PR-merge increase figures, and case details.