By Cursor shipped Projects in beta on September 10, and it’s worth reading the announcement carefully. On the surface, it looks like another iteration on agentic coding. It isn’t. Projects introduces a coordinator agent that runs on a cloud machine, stays alive when your laptop closes, and fans out to thousands of parallel subagents to do the actual work. The coordinator doesn’t write a single line of code. It plans, delegates, and brings results back for you to review. That’s a different category of tool.
The Coordinator-Subagent Model
Here’s the architecture that matters: when you create a Project in Cursor, you’re not starting an agentic chat session. You’re instantiating a persistent coordinator agent in the cloud. That coordinator reads your codebase, maintains context across sessions — months of it, Cursor says — and spawns subagents to implement whatever needs implementing. The subagents run in parallel, across as many machines as the work demands. When something requires local testing, the coordinator spins up a local agent for that.
Compare this to what AI coding looked like six months ago: you opened a chat, described a task, the agent tried to do it, you nudged it back on track. You were the loop. With Projects, the coordinator is the loop. You describe the goal once, check the output when it’s done.
A Project runs on its own computer in the cloud, so closing your laptop doesn’t stop it.
Cursor, Introducing Projects
Three Ways Projects Runs Without You
Projects ships with three trigger modes that require no human kickoff once configured:
- Slack: Point the coordinator at a bug-report channel. When a message comes in, it delegates the fix to a subagent and opens a PR. You come back to a draft pull request, not an empty to-do item.
- Schedule: Run on a clock — hourly, daily, weekly. Practical uses include daily scans for missing test coverage on recently merged code, weekly summaries of what shipped, and recurring security audits against a rules file you define.
- PR watching: The coordinator follows your open pull requests. CI failed? It delegates a fix. A review bot left comments? It addresses them. You don’t ask; it just does it.
The underlying design is Cursor’s Automations system, which launched earlier this year, now wired into the coordinator model for the Projects context.
The Numbers, and What They Actually Mean
Cursor claims users who primarily use Projects merge six times as many pull requests. New users in general are up 30%. The company didn’t publish cohort size, baselines, or time windows, so treat these as directional signals rather than benchmarks. The more important implication isn’t the 6x figure — it’s what 6x implies for review. If agents are opening six times as many PRs, the bottleneck for your team has moved from writing code to reviewing it. The constraint is now human attention, not implementation throughput.
Why the Coordinator Doesn’t Push to Main
There’s a deliberate design choice buried in the Projects announcement: the coordinator brings work back to you to approve. It does not push autonomously to main. This isn’t a technical limitation — it’s a trust decision. Fully autonomous agents that acted without checkpoints burned developer trust in 2024 and haven’t recovered it. Cursor watched that play out and built the review step in from the start.
That said, the tool is in beta. As with any agentic system, the quality of what comes back to you scales with the quality of context you give the coordinator upfront. Vague goals produce vague PRs. Treat the initial context you give each Project the same way you’d treat an onboarding brief for a new engineer.
What to Do Before You Enable It
Projects is in early beta on all paid Cursor plans starting at $20/month. Before you spin up a coordinator against a production codebase, two things to address:
- Set spend limits. Cursor rebuilt spend alerting in June 2026. Go to Settings → Billing and set dollar thresholds before enabling Projects. Tab completion is free; coordinator cloud machines and subagents are not. Parallel agents at scale bill fast.
- Start with a non-critical project. The feature is in beta and still evolving. Run it on a greenfield service or a backlog migration before pointing it at your main product.
The full changelog entry covers setup steps. Access is rolling out to all paid plans now. The era of AI pair-programming isn’t over — but the era of AI delegation has clearly started.
