Michael Truell's February thesis becomes a beta that remembers months of work, watches Slack and PRs, and keeps coding after the laptop closes.
By [RuntimeWire Staff](/author/runtimewire-staff)
· Published
Primary source: [Cursor on X](https://x.com/cursor_ai/status/2098162488013455784)
Why it matters #
Projects moves Cursor from helping with individual coding tasks to managing persistent software work. The strategic prize is control of the layer that assigns, remembers and reviews work across many agents.
Michael Truell (@mntruell) and fellow Cursor co-founders Aman Sanger (@amanrsanger), Sualeh Asif (@sualehasif996) and Arvid Lunnemark (@ArVID220u) launched Projects on September 10th, giving developers a persistent coordinator agent that directs coding work across cloud and local machines. Cursor announced the beta in a thread on X and detailed a system built for features, migrations and maintenance jobs that outlive a single chat.
The four founders met at MIT and started Anysphere in 2022. When Cursor raised its 2023 seed round, Truell described the goal as making programming "an order of magnitude faster" while Sanger argued that autocomplete left larger problems, including debugging and codebase navigation, open for a new product. Projects is their clearest move beyond the editor session: Cursor wants developers to supervise bodies of work while agents handle the implementation beneath them.
Truell laid out the product logic on February 26th in his "third era" thesis. He wrote that AI coding was progressing from autocomplete, through synchronous agents, toward fleets that operate for longer periods with less direction. Projects packages that thesis into a coordinator that developers can talk to, interrupt and redirect without waiting for the individual agents doing the work.
A manager for the agents
The coordinator does not ordinarily write code. According to Cursor's Projects launch post, it plans the work, creates subagents for research, implementation and testing, and collects their output for review. Cursor says one coordinator can delegate to thousands of subagents, though the launch materials do not specify account limits or the compute costs attached to that scale.
Each Project runs on a dedicated cloud computer, allowing work to continue after a developer closes a laptop. When a change needs to be tested against a local setup, the coordinator can start an agent on the developer's machine. Project files synchronize across those environments, carrying plans, demos, research, testing instructions and lessons about the codebase from one agent to the next.
That shared context addresses a familiar weakness in coding assistants. A new chat often means restating the architecture, conventions and decisions that shaped earlier work. Cursor's approach turns those lessons into durable project files that future agents can use. The useful part is cumulative: an agent that discovers how to test a service can leave instructions for every agent that follows.
Projects can also initiate work from outside signals. Cursor calls these connections "subscriptions." A coordinator can watch a Slack channel for bug reports, follow pull requests and react to CI failures, or run a task on a schedule. Cursor said the beta began rolling out to all users on September 10th through the left-hand navigation in its editor.
Cursor is testing the pitch on itself
Cursor says its engineers have used Projects for several months, including migrations spanning hundreds of pull requests, design-system maintenance and the development of Projects itself. One internal design-system Project scans new pull requests, extracts reusable components and creates a lint rule after encountering the same mistake twice. Cursor expects that workflow to touch 20 to 100 pull requests per day, with an engineer reviewing its work.
The internal deployment gives Cursor concrete examples beyond a staged demo. Its productivity figures need more scrutiny. Cursor says new Projects users merge 30% more pull requests and users who primarily work through Projects merge six times as many. Cursor supplies no sample size, measurement window or baseline for either comparison. The heavier-use group is defined by its adoption of Projects, making the 6x figure a product cohort metric rather than evidence that Projects caused the increase.
A 2025 randomized study from METR provides a useful counterweight to vendor productivity claims. Sixteen experienced open-source developers completed 246 tasks with AI tools allowed or prohibited; allowing the early-2025 tools increased completion time by an estimated 19%. Cursor was among the tools used. The study had a small sample, predates Projects and did not test persistent fleets working asynchronously, so its result does not settle the case against Cursor's new architecture. It does show why merged-pull-request counts cannot stand alone as proof of engineering productivity or quality.
Projects changes the unit Cursor is trying to optimize. Earlier assistants helped a developer finish a function, file or ticket. The coordinator is supposed to retain the reasoning behind a feature, continue maintaining it after release and respond when operational signals produce new work. That longer loop may matter more than raw code-generation speed if the agents can preserve intent and produce changes that humans can review without reconstructing every step.
Persistence is becoming table stakes
Cursor does not own the idea of a coding agent with a memory and a calendar. GitHub Copilot offers repository and user-level memory, along with scheduled automations. Replit Agent can split plans into background tasks that run in isolated project copies.
Cursor's specific bet is that a coordinator inside its development environment can unify those pieces: the codebase, editor, cloud agents, local testing, Slack signals, pull requests and recurring work. That integration also increases the blast radius of a bad instruction or a coordinator that learns the wrong convention. Cursor's own examples keep a human at the review step, even after the system has earned enough trust to operate across dozens of pull requests a day.
The launch arrives less than a month after Cursor said SpaceX completed its acquisition on August 14th. Cursor linked that transaction to access to more GPU capacity and lower model costs, without publishing the deal terms in the announcement. Projects offers a direct outlet for that compute: a persistent coordinator dispatching large numbers of cloud agents consumes inference as ongoing infrastructure rather than as an occasional chat response.
Truell's founding promise centered on speed and more time for creative work. Projects expands the wager. Cursor is asking developers to hand a growing share of planning, delegation and maintenance to a durable software manager, then judge the work through artifacts and pull requests. The product will earn that authority one reviewed change at a time.