Image: Cursor / cursor.com Cursor Launches Projects: One Coordinator Agent, Thousands of Subagents
Cursor's new Projects feature gives you a persistent coordinator agent that delegates coding work to thousands of subagents in parallel. Early data shows heavy users merging six times more pull requests.
Cursor shipped Projects on September 10, 2026, a new way to tackle large-scale coding work using a persistent coordinator agent that orchestrates other agents rather than writing code itself.
The idea is straightforward: instead of opening a fresh chat window for every task, you talk to one coordinator that stays open for the life of the project. That coordinator plans the work, breaks it into pieces, and hands each piece to a subagent with a clean copy of the codebase. When the subagent finishes, the coordinator brings the result back to you.
How it works
Three things make Projects different from a regular Cursor session.
Cloud execution. Each project runs on its own dedicated machine in the cloud. You can close your laptop and the work keeps going. Subagents spin up in isolated environments, each with fresh context, so they don’t step on each other.
Shared context. The coordinator maintains a set of files that sync across every cloud and local machine its agents use. This is what lets the system accumulate knowledge about your codebase over time. You don’t re-explain architecture decisions every session.
Subscriptions. This is the part that pushes Projects beyond a long-running chat. You can tell a coordinator to watch a Slack channel, follow your pull requests, respond to CI failures, or run on a schedule. The coordinator acts on what it notices, without you prompting it.
Projects launched in beta with a gradual rollout to all users starting September 10. You can start one from the navigation sidebar.
What Cursor says it’s good for
The company identifies three types of work where the coordinator model fits well.
Feature development. A coordinator can research your existing systems, delegate implementation to multiple agents working in parallel, and monitor the feature after it ships. Cursor describes this as enabling agents to “carry features from the first line of code to a merged PR.”
Migrations. Framework upgrades or system-wide refactors often span hundreds of pull requests. A coordinator can draft a migration plan, run agents in parallel, and validate incrementally. Each failed test is a signal the coordinator uses to adjust.
Ongoing maintenance. Design-system consistency, dependency updates, regression monitoring. These are tasks that nobody schedules but everyone wants done. With subscriptions, a coordinator can handle them continuously.
The numbers
Cursor cites two figures from internal testing. New users who try Projects merge 30% more pull requests than those who don’t. Users who primarily work through Projects merge six times as many.
The 6x figure comes with a note: it reflects users who are already bought into the workflow. The productivity gain from Projects scales with how much you use it.
The tradeoffs
The coordinator pattern amplifies both good work and mistakes. A small misunderstanding at the planning stage can propagate across dozens of subagent runs before anyone notices. The release notes from Cursor are honest about this, recommending strong CI, solid code review practices, and good observability as prerequisites before leaning on Projects for large migrations.
For quick, well-defined tasks, the overhead of setting up a coordinator probably isn’t worth it. Projects makes the most sense for work that takes days or weeks, not hours.
Context
Cursor announced $500M+ ARR earlier this year and has been moving aggressively toward autonomous agents. Self-hosted machines, cloud agents, and now Projects represent a clear direction: Cursor wants to be where long-horizon coding work happens, not just where you edit files.
The coordinator-and-subagent architecture is not unique to Cursor. Anthropic’s Claude Code has agent teams, and Devin’s whole pitch has always been a long-running coding agent. What’s new here is the combination of persistent context, subscriptions, and a clean interface for managing the coordinator through a regular chat-like thread.
Projects is in beta as of today. Pricing details for cloud compute costs have not been announced publicly.