I spend a lot of my days in Cursor with cloud agents and local subagents. The pain was never “can an agent write code?” — it was herding them. New chat, re-explain the repo, lose the thread, reopen the laptop to keep something alive. Cursor’s Projects beta (Sep 10) is the first thing that feels like it was built for that exact mess.
One coordinator, not a tab farm
A Project sits above the chat. You talk to a coordinator that doesn’t write the code itself — it plans, delegates to subagents, and brings finished work back for review. Cloud by default: the Project runs on its own machine, so closing the laptop doesn’t kill the fleet. When something needs to touch your box — a local test, a GUI check, a device only you have — the coordinator spins a local agent there.
That split matches how I already work: heavy lifting in the cloud, sticky local verification on the machine that owns the secrets, the simulators, or the weird network path.
Shared context that survives months
The useful part isn’t “more agents.” It’s that each Project keeps a shared file set synced across every cloud and local machine those agents use. Research, test recipes, preference notes — once one agent figures out how to hit a service, the next ones inherit it. Cursor’s own numbers are aggressive (Projects-heavy users merging far more PRs), but the product claim that matters day-to-day is simpler: stop re-onboarding the same codebase every Monday.
Subscriptions: PRs, schedules, team chat
From the changelog: point the coordinator at your PRs, a schedule, or a team chat channel (Slack, for shops that use it). It can fix CI, react when PRs open or merge, and pick up bug reports without you pasting the thread into a fresh agent. That’s the difference between “I run agents” and “something watches the workstream while I’m in a meeting.”
Cursor describes three patterns that map cleanly onto real engineering: feature work (research → plan → parallel implement/test), migrations (safe approach, then hundreds of incremental PRs), and gardening (design-system drift, recurring lint, endless quality work). The gardening example — a Project touching dozens of PRs a day while a human spot-checks — is basically how mature teams already want agents to behave.
Why this sticks for me
I’m not dropping Codex or local models. Cursor is probably my heaviest driver though. Projects is the layer that stops me from being the orchestrator of last resort. Start from the left-hand nav, describe a body of work that will outlive one chat, and let the coordinator own the herd. Beta, rolling out — but if your week looks like mine, this is the feature that makes multi-agent Cursor feel like a daily driver instead of a demo.
Sources: Introducing Projects · Cursor Projects changelog