Cursor Stopped Selling an Editor and Started Selling a Manager

On September 10, Cursor shipped Projects in beta. The feature is rolling out to all users. It exists for work that is too big for a single prompt, which Cursor describes as "a feature, a migration, or a full app."

The key design choice is who does the work. Each Project gets a coordinator agent, and by Cursor's own description the coordinator "doesn't write code itself." Instead it plans the work, splits it up, hands the pieces to agents that write the code, and returns the finished result for a human to check. Cursor says it can delegate to thousands of subagents, keep context over months of work, and run on its own cloud computer, so "closing your laptop doesn't stop it."

This is the first time a major coding tool has put a manager at the center of the product instead of a coder. That matters more than the subagent count.

What actually shipped

Cursor's changelog makes three concrete claims.

Persistence. A Project runs in the cloud, not on your machine. Files sync across every agent working in it. Agents add research, artifacts, and what they learn about the codebase to a shared context, so the Project builds up knowledge instead of starting from zero each session.

Parallelism. The coordinator creates and manages agents for you, "running as many in parallel as the work needs." Cursor gives no hard limit. Its headline says the coordinator can delegate to thousands of subagents.

Triggers. You don't have to prompt a Project. You can tell the coordinator to watch a Slack channel, run on a schedule, or follow all your pull requests, and it delegates work whenever a matching signal comes in. Cursor's example is a bug-report channel: each new report becomes delegated work without anyone typing a request.

The third claim is the one that changes the category. A coding assistant waits for you. A Project with triggers is closer to a standing team that watches your inbox.

The unit of work just changed

For three years, AI coding products have been priced and judged per interaction: one autocomplete, one chat turn, one agent run. Everything, from the benchmarks to the demos to how developers talk about these tools, assumed a human starts each piece of work.

Projects breaks that assumption in two ways. The unit becomes a body of work that lasts months. And the person starting it becomes a signal, like a Slack message, a failing CI run, or a cron schedule, not a developer at a keyboard.

That changes what developers spend their time on. If the coordinator plans and subagents write, the human job is scoping and review. Scoping is a skill most engineers are good at. Reviewing the output of a thousand parallel workers is a job nobody has done before. The limit on a Project won't be how much code the agents can write. It will be how much of that code a person can responsibly approve. Cursor's changelog says the coordinator brings work back "for you to check," but it doesn't say how that review scales once the number of subagents gets large.

The economics point the same way

Cursor's pricing page lists cloud agents on every paid tier: Pro at $20 a month, Pro+ at $60, Ultra at $200. Each plan includes a set amount of model usage. Beyond that, on-demand usage is "billed in arrears." The Projects changelog does not list a separate price.

Put those facts together and you get the business model. A developer using an editor spends compute at the speed they can type. A coordinator spreading work across subagents, starting itself from Slack and CI, spends compute at the speed of the backlog. For a vendor with usage-based billing, Projects makes revenue scale with how many agents are running, not with how many seats a customer buys.

That also explains the timing. SpaceX closed its $60 billion all-stock acquisition of Cursor in mid-August. Cursor's announcement of the deal promised access to "the largest fleet of GPUs in the world." A product whose default behavior is massive parallel cloud execution is the most direct way to put that fleet to work. Projects is what an AI coding company builds after its owner hands it a data center.

Fewer models behind more agents

The same month exposed the other side of the deal. On August 28, OpenAI said it would cut off Cursor's access to its models on November 12. OpenAI said it cannot be confident SpaceX will use its technology within its terms of service, and it will not supply its upcoming Astra model to Cursor at all. According to DevOps.com, Cursor CEO Michael Truell said OpenAI models account for about 5% of Cursor's user traffic.

Cursor's pricing page now advertises "generous limits for Grok" on every paid plan. Anthropic's Claude and Google's Gemini models are still available. But the list of models is getting shorter at the same moment Cursor is asking customers to trust it with months-long, largely unsupervised runs.

That is the tension in Projects. A coordinator is only as good as the models it hands work to. Thousands of parallel subagents make any weakness in those models worse, whether the weakness is cost, speed, or how often they make things up. The company that owns Cursor also owns the model it promotes most heavily.

What to watch

Review tooling. The first thing Cursor has to ship after Projects is a better way to approve what Projects produces. Without it, the parallelism mostly creates a bigger review queue.

Spend controls. Signal-triggered agents billed in arrears can surprise a finance team. Watch for per-Project budgets and caps on how many agents can run at once.

Model mix after November 12. Once OpenAI's models are gone, it will show whether Cursor stays a neutral tool for choosing models or becomes the developer storefront for Grok.

Whether competitors copy the coordinator. The main idea here, a manager agent that never writes code, is easy to describe and hard to do well. If rival coding tools announce their own versions this quarter, Cursor will have set the pattern. If they don't, it will be because long-running delegation turned out to be harder to sell than it is to demo.