What does the coordinator agent in Cursor Projects do?

It directs rather than codes. Cursor states the coordinator 'doesn't write code itself but directs other agents that do,' which means it is never blocked by a long task and always responsive to your instructions. It creates and manages subagents, runs as many in parallel as the work needs, decides when something must be tested on your local machine and spins up a local agent for it, and — through subscriptions — can watch a Slack channel, run on a schedule or follow all your PRs, fixing CI and acting when they open or merge.

What is shared context in Cursor Projects?

A set of files each Project maintains and syncs across every cloud and local machine its agents use. Agents write research, artifacts, and what they learn about the codebase and your preferences into it — for example, once one agent works out how to test a service, every future agent can reuse those instructions. The context grows with the Project, so you do not re-onboard an agent for each task and the coordinator becomes more effective over time.

How much does Cursor Projects cost and is it generally available?

As of September 14, 2026 Projects is in beta and rolling out to all Cursor users; Cursor has not published Projects-specific pricing. Projects run on Cursor's Cloud Agents, so usage is billed through your Cursor plan, which starts at $20 per month for Pro. Because a Project runs on its own cloud computer and can run thousands of subagents, expect usage to scale with the amount of parallel work you authorize, and review the coordinator's plan before approving large fan-outs.

Quick Answer

What Is Cursor Projects? Coordinator Agents Explained

Q: What is Cursor Projects?

Cursor Projects is a feature of the Cursor AI code editor, launched in beta on September 10, 2026 and rolling out to all users, for bodies of work larger than a single chat — a feature, a migration, or a full app. You start a Project from the left-hand nav, describe what you want, and a coordinator agent takes over: it plans the work, delegates to as many cloud subagents as needed, keeps a shared-context file set that syncs across machines, and brings finished pull requests back for you to review. Cursor says it maintains context over months and can perform recurring work without being prompted.

Published: September 14, 2026

The short answer

Cursor Projects moves you one level up from prompting an agent to directing a coordinator that runs agents for you. Launched in beta on September 10, 2026, a Project has its own cloud computer, a coordinator that plans and delegates but never writes code, a shared-context store that persists across months, and subscriptions that let it act on Slack messages, schedules and PRs without being asked.

	Cursor Projects (as of September 14, 2026)
Launched	September 10, 2026, beta, rolling out to all users
Where	Left-hand nav in Cursor; describe the work, coordinator takes it from there
Coordinator	Plans, delegates to subagents, reviews, reports back — “doesn’t write code itself”
Scale	”Delegates tasks to thousands of subagents,” as many in parallel as the work needs
Compute	Cloud by default (Project has its own computer); local agent spun up when something must run on your machine
Memory	Shared-context files synced across every cloud and local machine the Project’s agents use
Triggers	Subscriptions: Slack channel, schedule, follow all PRs / fix CI
Cursor’s results	New users merge 30% more PRs; Projects-primary users merge 6x as many
Pricing	Not published; billed as Cloud Agent usage on Cursor plans (Pro from $20/mo)

Why Cursor built it

In February 2026 Cursor described a “third era of software development” in which fleets of agents take on entire bodies of work. Projects is its implementation. The company says it used Projects internally for several months before launch — running migrations of a few hundred PRs, keeping its design system consistent, and building Projects itself — and frames the benefit as freeing developers “from managing agents” so they can “direct the work itself.”

The problem it targets is real: single-chat agents lose context between sessions, cannot run more subagents than your laptop supports, stop when you close the lid, and need re-onboarding for every task. Projects attacks each of those directly.

The three core capabilities

1. Cloud by default, local when needed

A Project runs on its own computer in the cloud. That gives it two properties a laptop agent cannot have: it keeps working when you disconnect, and it can run more subagents in parallel than your machine could host. When a step needs your environment — a device test, a local database, a build that only works on your Mac — the coordinator spins up a local agent to run it there and reports back. Cursor’s separate self-hosted machines feature (expanded in early September 2026) lets organizations run agent tool calls inside their own network, keeping code, build outputs and secrets on internal hardware, with computer use on Linux and Mac.

2. Shared context

Every Project maintains a set of files that sync across all the cloud and local machines its agents use. Agents contribute research, artifacts, codebase insights and learned preferences. Cursor’s example: if one agent figures out how to test a service, every future agent can use those instructions. The store grows with the Project and makes the coordinator more effective the longer it runs — effectively a team memory that is written by agents rather than by you.

3. Subscriptions

Tell the coordinator to watch a Slack channel, run on a schedule, or follow all your PRs. It then acts on signals — a bug report lands in Slack and it delegates a fix; a PR opens and it fixes CI; a nightly run tidies the codebase — without waiting for a prompt. This is what turns a Project from a task runner into an always-on teammate.

How Cursor’s own engineers use it

Feature work. Agents research the system and record what they learn as shared context; the coordinator plans and sends agents to implement and test parts in parallel; when the feature is ready, it starts a local agent so you can try it; after shipping, the same Project can monitor logs and handle bug reports with full context of the original decisions.
Migrations. “Easy to start and difficult to finish.” You agree a safe approach with the coordinator; it applies it incrementally across hundreds of PRs; you review closely at first and less as fixes hold up.
Gardening. Never-ending maintenance. One engineer’s design-system Project follows every new PR, extracts components that belong in the design system, and adds a lint rule whenever it sees the same mistake twice — on track to touch 20 to 100 PRs a day.

What Projects is not

Not an IDE feature for small edits. Cursor says it works best on “work that will outlive a single chat.” For a one-file fix, Agent mode is faster.
Not a model. The coordinator and subagents use Cursor’s model menu. Note that Cursor’s access to OpenAI models ends November 12, 2026, so a Project’s model pool after that is Anthropic, xAI, Cursor’s Composer line and open-weight models.
Not free of review. The coordinator “brings the finished work back to you to check.” Cursor’s own migration story is review-heavy early and lighter later — autonomy is earned, not default.
Not priced yet. Beta, with usage flowing through Cloud Agent billing. Large fan-outs deserve a cost check before approval.

How it compares

Cursor Projects is a product; OpenAI’s Agents API (also September 10, 2026) is the platform for building the same pattern on the Codex harness; Claude Code is the agent most teams use interactively, with Cowork remote sessions covering the “keep working when I close my laptop” case. If you want the coordinator today and live in Cursor, Projects is the shortest path. If you need OpenAI models in a coordinator after November, or your own policy layer, build on the Agents API instead.

Getting started checklist

Open Projects in the left-hand nav and describe a body of work that will outlive one chat — a migration is the canonical first Project.
Let agents research first; check the shared-context files they write so the coordinator’s plan rests on correct assumptions.
Review every PR early. Loosen review only after a run of clean merges.
Add a subscription (PRs or a Slack channel) once you trust the approach.
Watch usage: a Project’s cloud computer and parallel subagents bill against your plan.

Last verified: September 14, 2026.