The Project becomes the persistent unit of agentic engineering
Source → Observed → Interpretation → Model implication
Introducing Projects
View source →Cursor describes Projects as maintaining context over months, delegating tasks to thousands of subagents, and performing recurring work without a new prompt.
Cursor reports using Projects internally for several months for features, migrations spanning a few hundred pull requests, design-system maintenance, and development of Projects itself.
Cursor reports that new Projects users merge 30% more pull requests and users who primarily use Projects merge six times as many.
Each Project has a coordinator agent that plans and delegates but does not write code; implementation and testing are assigned to other agents that can operate in parallel.
A Project runs on its own cloud computer and can continue when the developer's laptop is closed, while the coordinator can start a local agent when work requires testing on the developer's machine.
Project files synchronize across the cloud and local machines used by its agents. Agents add research, artifacts, codebase knowledge, and user preferences so discoveries made by one agent are available to future agents.
A coordinator can watch Slack, run on a schedule, or follow pull-request events and act on detected signals without waiting for a new user prompt.
Cursor reports that its migrations establish a safe approach with the operator, apply it incrementally across hundreds of pull requests, and continue with less review as fixes hold up.
Cursor reports one design-system Project that scans new pull requests, extracts reusable components, and adds a lint rule after seeing the same mistake twice. Cursor says the Project is on track to touch 20 to 100 pull requests per day.
Projects extends Cursor's differentiated planner-worker architecture from a bounded swarm experiment into persistent production work. Context, cloud and local execution, event observation, accumulated learning, and human review are integrated around a durable Project rather than reconstructed for each agent session. The coordinator and worker agents have stable different responsibilities, while shared context and subscriptions let their cooperation persist across machines, tasks, and time. Cursor's reported pull-request differences provide observational support for a cooperation advantage, although they do not isolate causality. More strongly, the Project itself is the object engineers create, direct, revisit, and hold responsible for a body of work, supporting the model's higher-order Specialization claim.
SUPPORTS. Supports Specialization by documenting persistent differentiated cooperation in ordinary engineering work: coordinators plan and delegate, workers implement and test, cloud and local environments provide distinct execution capabilities, and shared project context preserves learning across executions. These parts operate together over months and across hundreds of pull requests, with the Project becoming the meaningful unit engineers direct. Reported pull-request differences also support Cooperation, but remain observational and do not establish the causal contribution of individual mechanisms.
What this does not establish
- Delegation to thousands of subagents describes work across a Project and does not establish thousands of agents operating concurrently.
- Cursor does not publish the definitions, denominators, time windows, quality controls, or statistical methods behind the 30% and six-times pull-request comparisons.
- Users who primarily use Projects may differ from other users, so the reported pull-request differences do not isolate Projects as the cause.
- Pull-request volume does not by itself establish equivalent improvements in software quality, customer value, cost, reliability, or review burden.
- The design-system Project is reported as being on track to touch 20 to 100 pull requests per day; the source does not establish sustained operation at that rate.
- Cursor's internal results and product claims are first-party and have not been independently replicated across other organizations or engineering environments.
Across repeated production engineering work, do persistent Projects preserve quality and reduce human coordination and review cost as their shared context, subagent population, event sources, and pull-request volume grow?