Cursor Projects Uses a Persistent Coordinator Agent to Manage Months-Long Development and Can Launch Thousands of Subagents Based on Events

Cursor launched the beta version of Projects on September 10. Its central innovation is not another agent that writes code, but a persistent coordinator placed between developers and execution agents. In a single ongoing conversation, users describe a feature, migration, or maintenance objective. The coordinator handles research, decomposes the work, delegates implementation and testing, and returns the results for review, but does not modify code directly.

Each Project includes a cloud execution environment that continues running after the developer closes their laptop. A local agent is launched only for tasks that must be validated on the developer’s device. Research notes, plans, artifacts, and team feedback are stored in shared files and synchronized across the machines used by the agents. This is intended to address a recurring problem in multi-agent workflows: agents repeatedly having to relearn the codebase and duplicating the same investigations. Subscriptions can monitor Slack channels, schedules, and PR open or merge events, allowing tasks such as fixing CI, handling bug reports, or cleaning up code to begin without waiting for a new prompt.

Cursor says the system can orchestrate thousands of subagents and has also used it internally for framework and design-system migrations spanning hundreds of PRs. One design-system project is expected to process 20 to 100 PRs per day. The company also says new users merge 30% more PRs and that heavy users merge six times as many. However, these figures are internal correlational data that do not control for the type of work, team size, or self-selection bias, so they should not be treated as direct evidence from a productivity experiment.

Engineering teams should pay closer attention to the as-yet-undisclosed costs of parallel execution, failure recovery, conflict resolution, and the permission model. A persistent agent that can read Slack, write to repositories, and run tests locally effectively resembles a service account with long-lived credentials. Deployments should therefore continue to restrict token scopes, retain branch protection and human approval gates for merges, and begin with a small migration to measure the model, compute, and review costs of each successfully merged PR.