AI Programming Tool Cursor Announces Launch of Multitask Capability in New Cursor 3 Interface
AI programming tool Cursor has announced the launch of the "/multitask" capability in the new Cursor 3 interface, supporting parallel processing of requests through asynchronous subagents: when users issue multiple tasks, Cursor no longer queues them serially but instead starts parallel subagents to run in the background for different tasks, allowing the main session to continue working. For messages that have already entered the queue but have not yet been executed, users can also directly invoke the /multitask command to convert these requests from "waiting in queue" to "executing in parallel," reducing the issue of long-chain tasks occupying front-end interaction time.
Source: Public Information
ABAB AI Insight
The introduction of multitask essentially upgrades the linear interaction model of "one large model doing one thing at a time" to a multithreaded execution model where "a master agent schedules multiple subagents to work in parallel." Previously, Cursor's subagents operated mainly in a synchronous manner, requiring the master agent to wait for sub-tasks to complete before continuing. Now, asynchronous subagents can complete tasks like code searching, testing, and refactoring in the background while the main dialogue continues to accept new instructions or review other changes, effectively embedding a self-scheduling "AI multiprocess system" within the IDE.
From the perspective of development process structure, this parallelization changes the location of the "time bottleneck": developers previously often had to wait for a single long task (such as cross-file refactoring or large-scale testing) to complete. Now, these tasks can be handled by background subagents, allowing developers to advance other issues in the foreground or assign different dimensions of testing, documentation, and refactoring to multiple subagents. This means AI is not just "a smarter autocomplete" but begins to act as "multiple junior engineers in the engineering team," with developers serving as coordinators and reviewers to allocate tasks.
On a deeper level, /multitask pushes AI tools from being "single-threaded assistants" to "task orchestration platforms," which will have long-term impacts on software productivity and organizational structure. The real constraints will gradually shift from "how powerful the model is" to "how tasks are divided, how parallelism is designed, and how results are merged and reviewed," transitioning from competition based on model capabilities to competition based on "AI engineering and process design." In this framework, teams capable of defining subagent roles, task segmentation, and quality thresholds will gain structural advantages in delivery speed and reliability compared to teams that rely solely on a large model.