Cursor rolled out Projects to all users in beta, a feature that lets a single coordinator agent run a body of work, a feature build, a framework migration, an entire app, across weeks or months rather than a single chat session. The coordinator does not write code itself. It delegates to subagents that do, then keeps working after you close the laptop.
That persistence is the structural bet. Cursor built Projects to run cloud-first, spinning up a local agent only when something needs to be tested on your own machine. Each Project also keeps a shared context file that syncs across every agent that touches it, so a lesson one subagent learns (how to test a particular service, say) becomes available to every subagent that follows.
Cursor says Projects has already changed how its own engineers ship. The company reports it has used the tool internally for months to run migrations spanning hundreds of pull requests and to keep its design system consistent while building Projects itself.
The company’s headline number: new users who adopt Projects merge 30 percent more pull requests, and engineers who primarily rely on it merge six times as many. Both figures come from Cursor measuring its own user base, not from an independent audit. A rise in merged pull requests measures throughput, not the quality of what shipped, whether it required more review cycles to land, or whether the resulting code held up in production. Cursor’s announcement includes no data on bug rates, revert frequency, or review time saved.
The “thousands of subagents” framing is also Cursor’s own description of scale, offered without a count of concurrent agents in a typical Project or a cost breakdown for running that many cloud instances. For a team evaluating whether to adopt this, the open question is not whether agents can merge code faster. It is whether a coordinator model, reviewed lightly after the first few pull requests, catches the failure modes that only show up after a migration has already touched production.
Cursor frames three recurring internal patterns for Projects: feature work, where a coordinator researches a system before implementing it in parallel; migrations, where a team establishes a safe pattern once and then applies it incrementally with less oversight over time; and what it calls gardening, ongoing maintenance where a coordinator watches Slack channels or pull request activity and acts on triggers without being prompted. One internal example Cursor cites: a design-system Project that started with an engineer correcting every fix and is now positioned to touch between 20 and 100 pull requests a day with lighter review.
That gardening pattern is the most notable claim in the announcement, because it describes an agent acting on a schedule or a detected signal rather than a direct prompt. Autonomous, subscription-style triggering is where coding agents have historically been weakest: false positives compound quickly across dozens of daily pull requests if the coordinator misjudges even a small fraction of them. Cursor’s post does not disclose an error rate for the design-system example, only that review intensity decreased over time.
Cursor is positioning Projects against a broader thesis the company laid out in February: that software development is entering an era where developers direct fleets of agents rather than manage them task by task. Whether that abstraction holds depends less on the coordinator’s planning and more on how well its shared-context system prevents subagents from repeating the same mistakes across a codebase, since that is the one internal claim Cursor makes that is checkable by any team that adopts the product.
Teams running long migrations or maintenance backlogs should pilot Projects on a bounded, reversible task first, track defect and revert rates against the pull-request count, and only extend the coordinator’s autonomy on gardening-style work after that data, not Cursor’s own PR-volume figures, clears the bar.
Cursor announced Projects in a September 2026 company blog post titled “Introducing Projects,” bylined Alexi Robbins and Fredrika Lindh.
