Anysphere launches Cursor 3, an AI code editor with local and cloud agents

Anysphere unveiled Cursor 3, a new version of its AI programming tool, on April 2. The company is rebuilding the product around an agent mode: Cursor is now meant not just to suggest code but to coordinate several assistants that write, check and refine changes.

A new Cursor interface

The main change in Cursor 3 is a new interface built from scratch around AI agents rather than the classic editor window. A developer can describe the needed function in plain language, choose one of the supported models, including Claude, and get not only generated code but also a demonstration of how that code works. In effect, Cursor is moving further away from the role of smart autocomplete toward that of a control panel where a person manages a whole set of automated workers.

"Cursor 3 is a unified workspace for building software with agents."

The new window gathers all active assistants: local ones, cloud ones, and those launched from other channels. Each agent has its own working context, and the user sees a shared task list in the sidebar. This matters for teams and developers who work across several repositories at once: Cursor 3 is designed for parallel work, where one agent writes code, a second prepares edits, and a third shows results through screenshots and demos.

Local and cloud agents

Anysphere has split the work between two types of agents. Cloud agents get more computing resources, so they suit heavy, long-running tasks, such as generating large chunks of code or processing several requests in parallel. Local agents are slower, but they have a different advantage: the developer can open the result immediately on their own computer, make manual changes, run tests and quickly check how everything behaves in a real environment.

An important detail of the release is fast session transfer between the cloud and a local machine. You can start a task in the cloud and then hand it to the desktop for fine-tuning. Or the reverse: send a local session to the cloud so the agent keeps working while the user goes offline or switches to another task. For these scenarios Anysphere singles out Composer 2, its own programming model, which the company calls more cost-efficient for such workflows.

What's in the release

Cursor 3 adds not one big feature but several layers of convenience around agent-based development. Some target code generation and editing; others target quality control and speeding up the everyday cycle from request to finished change. Besides the agents themselves, Anysphere updated the interfaces for viewing diffs, reviewing edits and navigating tasks, so developers can more easily not only launch the AI but also accept its work.

• a chat interface where a task can be described in natural language
• switching between local and cloud agents without breaking the session
• a design mode for quickly editing interface elements via text instructions
• a step-by-step view of the agent's actions with explanations, errors and screenshots
• faster review of changes before a commit and a production release

Anysphere also improved the answer-selection scenario: a developer can now send a request to several large language models at once and pick the best result. This reduces dependence on a single model and makes Cursor more of a shell for working with different AIs than just another editor with a built-in chatbot. In practice, this approach helps when you need to compare implementation options quickly rather than blindly accept the first generated answer.

What it means

Cursor 3 shows how quickly the market for AI programming tools is moving from the "in-editor suggestion" format to an "agent orchestrator" format. Anysphere is betting that developers will increasingly not write every line themselves but distribute tasks among local and cloud assistants while checking intermediate results. If this scenario catches on, the next field of competition will be not individual models but how easy it is to manage a whole team of agents.