Cursor launches Projects in beta: cloud coordinator dispatches thousands of parallel subagents across month-long workloads
Engineers who used Cursor Projects as their primary workflow merged six times as many pull requests as those who did not, per Cursor at the September 10 launch of the feature. New users who ran Projects alongside the standard IDE merged 30% more PRs. Projects places a coordinator agent in charge of an entire body of work and delegates to thousands of parallel subagents running in cloud-hosted machines, with no requirement that the developer's laptop stay open, per the Cursor launch post.
What launched
Cursor shipped Projects in beta on September 10, 2026, written up by Alexi Robbins and Fredrika Lindh in the official announcement. The feature is accessible from the left-hand navigation panel in the Cursor IDE. Users describe a body of work (a feature, a migration, or a full application), and the coordinator agent builds a plan, assigns tasks to subagents, and manages the results back to the developer.
The coordinator does not write code. It plans and delegates to other agents that do, which keeps it responsive to direction at any point in a long-running session. If the developer wants to redirect the work or weigh in on a decision, they do not need to wait for a sub-task to finish.
Cursor has been using Projects internally for several months before the launch, per the launch post. The team ran migrations across hundreds of pull requests, maintained design-system consistency, and used the tool to ship Projects itself.
Three pillars of the architecture
Cursor describes three capabilities that separate Projects from a standard agent session.
Cloud by default, local when needed. A Project runs on a cloud-hosted machine provisioned for the task. Closing the developer's laptop stops nothing. When the coordinator needs to test something on the developer's local environment, it spins up a local agent for that step and returns work to the cloud when done.
Shared context. Each Project maintains a set of files that synchronize across every cloud and local machine involved in the work. Agents write back what they learn: how the codebase is structured, how to reach a particular service, what test patterns the developer prefers. If one subagent learns how to run the integration suite, the next subagent starts with that knowledge already in place. Context accumulates over the life of the Project rather than resetting between sessions.
Subscriptions. A coordinator can watch a Slack channel, run on a schedule, or track every pull request in a repository. When a bug report lands in Slack or a PR opens, the coordinator dispatches subagents without waiting for a prompt. Cursor cited one engineer who runs a design-system Project this way: the coordinator scans each new PR, extracts reusable components, and adds a lint rule whenever it spots the same mistake twice. That Project is on pace to touch 20 to 100 PRs per day, per the launch post.
How Cursor uses it internally
Three patterns cover most of what Cursor's own engineers do with Projects.
Feature work typically starts with research agents mapping the codebase and recording what they find as shared context. The coordinator then sends agents to implement and test different parts in parallel. After the feature ships, the same Project can monitor logs and handle bug reports, carrying the original design decisions as context.
Migrations are where Cursor reports the clearest gains. The team adopted new frameworks and replaced styling systems across hundreds of PRs using Projects. A developer reviews each fix closely early on; as the fixes hold up under review, the coordinator continues the migration without waiting for manual prompts.
Maintenance work (regression checks, code-quality sweeps) maps to the subscriptions capability. The coordinator acts when a condition fires rather than running on a human-set clock.
Why it matters
The coordinator model shifts who owns the task queue. In a standard agent session, the developer manages context resets, sequences steps, and stays online while the agent runs. Projects moves that orchestration responsibility to the coordinator agent.
The shared context layer is the most consequential piece for teams running large codebases. The ramp cost of spinning up an agent on a mature codebase (reading docs, mapping dependencies, learning conventions) is often larger than the coding task itself. If shared context works as Cursor describes, that cost becomes a one-time investment per Project rather than a recurring fee per prompt.
The subscription capability extends the tool beyond interactive sessions into territory closer to CI/CD automation and async incident response. A Project tied to a bug-report Slack channel can respond to an incident without a developer at their desk. That is a distinct product category from an IDE copilot.
The 6x PR-merge-rate figure is per Cursor's own reporting and should be read as internal company data rather than an independent benchmark. But the mechanism behind it is legible: subagents working in parallel across cloud machines, accumulating shared context, and never blocking on the developer's availability.
Getting started
Projects are available in beta and rolling out to all Cursor users from September 10. To start, open Cursor (affiliate link), navigate to the left-hand panel in the IDE, describe the work, and the coordinator takes over. Cursor recommends Projects for bodies of work that will span multiple pull requests or that the developer wants to continue while away from the machine.
What to watch next
Cursor has not published the pricing or compute model for cloud-hosted project machines. The fee structure for long-running multi-subagent sessions will determine whether Projects is economically viable for large migrations or mainly for focused feature work.
OpenAI launched its Agents API on September 10, the same day Cursor shipped Projects. The immediate competitive overlap is narrow (OpenAI Agents API targets developers building their own agents; Cursor Projects targets software engineers using a finished product), but the convergence of launch dates marks the same week that cloud-orchestrated multi-agent coding crossed from research preview to generally available product.
First independent benchmarks comparing coordinator quality and cost across Cursor Projects, Codex-backed agent setups, and similar frameworks will be the next useful signal for teams deciding where to invest.
Sources
- Introducing Projects: Cursor blog, September 10, 2026, by Alexi Robbins and Fredrika Lindh (primary)
- Cursor Projects changelog: Cursor, September 10, 2026 (primary)