Cursor launches Projects to coordinate AI subagents

Cursor has released a beta feature called Projects, shifting the AI-assisted coding workflow from single-task chat sessions to a persistent, multi-agent orchestration system. Instead of interacting with individual agents, developers communicate with a single coordinator agent. This coordinator does not write code itself, allowing it to remain responsive while delegating parallel tasks to thousands of subagents running on cloud servers or local machines.

The system relies on three core capabilities: cloud-first execution, shared context, and subscriptions. Because Projects run in the cloud, they continue executing even when a developer closes their laptop, spinning up local agents only when machine-specific testing is required. A synchronized set of context files accumulates codebase knowledge over time, eliminating the need to repeatedly onboard agents. Additionally, subscriptions allow the coordinator to monitor Slack channels, track pull requests, fix continuous integration (CI) pipelines, or execute tasks on a schedule without manual prompts.

According to Cursor, the workflow significantly boosts developer output. Internal testing shows that new users merge 30 percent more pull requests, while heavy adopters of Projects merge six times as many. The tool is designed for substantial, long-term engineering efforts rather than quick, five-minute refactors. Ideal use cases include building new features, managing framework migrations, and routine codebase maintenance, such as running a design-system project that can touch 20 to 100 pull requests a day.

For software engineers, this development addresses the persistent issue of context rot, where developers must constantly re-explain system architecture in new chat windows. By treating accumulated knowledge as a shared, persistent asset, Projects transforms the IDE into an autonomous background worker. However, this automation introduces risks. Without rigorous code reviews and robust testing pipelines, minor errors can quickly compound across hundreds of automated pull requests if the shared context drifts or subagents misinterpret coding conventions.