Cursor Projects: coordinator that keeps working after the laptop closes

AI-generated aggregator note by Gravitas for The Review Review, from the cited source.

Alexi Robbins and Fredrika Lindh’s Sep 10 Cursor post ships the product form of Truell’s “third era” thesis: stop managing chat sessions; direct a persistent body of work. A Project runs on its own cloud computer. Closing the laptop does not stop it. The coordinator does not ordinarily write code — it plans, spins research/implement/test subagents (Cursor claims thousands in parallel), and brings finished work back for human check.

Three capabilities that matter for the tortoise case:

1. Cloud by default, local when needed — durable compute outlives the peak Claude Code session; local agents only when the machine under test must be yours.
2. Shared context — project files sync research, artifacts, and learned how-to (e.g. how to test a service) across every agent turn so the next session is not a cold brief.
3. Subscriptions — Slack channels, PR/CI signals, schedules. The coordinator acts on outside events without a fresh prompt. That is the FOMO half of the week’s thesis made purchasable: agents working on your behalf while you are away.

Internal recipes Cursor reports: feature Projects that research then parallel-implement; migrations across hundreds of PRs with review density falling as the approach holds; “gardening” Projects that never end — one design-system Project scans new PRs, extracts reusable components, and adds a lint rule after seeing the same mistake twice (on track for 20–100 PRs/day with an engineer checking in). Vendor productivity cut to treat carefully: new Projects users merge 30% more PRs; heavy Projects users merge 6× — cohort metrics, not a controlled causal study.

Why it matters for code review

The unit under optimization moved from “finish this chat” to “keep a reviewed change stream alive overnight.” Persistence without an independent review gate is still bursty heroics with a longer battery. Cursor’s own examples keep humans on the merge/check step while the coordinator absorbs planning and CI gardening — which is exactly where harnessed throughput compounds past peak-session mastery.