Cursor 3 Ditches VS Code Core for AI Agent Management
Cursor, once a popular AI-powered code editor built on Visual Studio Code, has unveiled a radical new direction with version 3.0. The latest release completely rewrites the editor from the ground up in Rust, moving away from its VS Code roots. Instead of focusing on traditional code writing, Cursor 3.0 aims to be a central hub for managing multiple AI agents that can work across different projects, machines, and even the cloud simultaneously.
This significant shift positions the tool more like an air traffic controller for AI, a stark contrast to its previous iterations. Cursor 1.0, released two years ago, offered AI code completion akin to a co-pilot. Six months later, Cursor 2.0 introduced an upgraded chat view capable of controlling the terminal to build features, acting more like a captain. Now, Cursor 3.0 wants users to orchestrate swarms of AI agents, abstracting away much of the direct coding process.
Introducing Composer 2: A New In-House AI Model
Accompanying the release of Cursor 3.0 is Composer 2, Cursor’s new in-house coding model. Initially, Cursor claimed Composer 2 outperformed leading models like Claude Opus 4.6 on internal benchmarks, boasting higher intelligence at a lower cost and faster speed. However, transparency concerns arose when it was discovered that Composer 2 is based on Moonshot’s Kimmy K2 model. This revelation came after a model ID was found in Composer’s metadata, leading to community discussion.
Adding to the complexity, the Kimmy model itself has faced accusations of being trained on outputs from other AI models, sometimes even identifying itself as Claude. Cursor has since apologized for its lack of transparency regarding Composer 2’s origins. Despite the initial confusion, the company released a detailed technical report on their development process, emphasizing that even an AI model that is smart, fast, and affordable will be highly valuable in a future where less direct coding is required.
A New Interface for Agent Orchestration
To facilitate this shift towards agent-based development, Cursor’s interface has been completely rebuilt using Rust and TypeScript. While the familiar VS Code editor remains accessible, the new interface prioritizes managing AI agents. Users can now largely delegate code generation to these agents, with Cursor acting as a central platform that combines the power of a professional development environment with language servers, file management, remote SSH capabilities, and the ability to run numerous agents concurrently.
This approach has drawn comparisons to OpenAI’s Codex, but Cursor 3.0 emphasizes its capabilities in managing multiple, parallel AI tasks. The system allows users to monitor the status of each agent, with visual cues indicating when human input is needed (yellow dot) or when a task is complete (blue dot). This allows for seamless multitasking across different projects or remote servers.
Real-World Application: Building a Project with AI Agents
Demonstrating the new capabilities, the developers showcased building a prototype for a fictional company called “Horse Tinder” using a swarm of AI agents. What might have taken months of manual coding or weeks of traditional AI-assisted coding was reportedly completed in a matter of minutes. The process began by entering “plan mode” for the AI to outline the project’s architecture.
While the main architecture was being developed, other agents were simultaneously tasked with creating a separate landing page for marketing and performing remote work on a cloud server. Users can switch between these tasks and even different projects without interruption. The system allows for parallel agent work, with progress easily monitored from a single window. This rapid prototyping capability highlights the potential for significant acceleration in software development.
Iterative Design and Refinement
Once the initial code was generated, Cursor 3.0 provided tools for immediate review and refinement. Developers could access the git history, open a terminal, and inspect files within a minimal file explorer. A built-in browser even allowed direct interaction with the developing application, enabling real-time testing of features like a rendered SVG graphic.
The design itself was critiqued as “horshit,” but the platform’s “design mode” allowed for quick, AI-driven modifications. Instead of manually adjusting CSS, users could highlight elements and instruct the AI to fix them, with changes occurring in the background. This iterative feedback loop, where AI handles the underlying code adjustments based on high-level design requests, further streamlines the development process.
Why This Matters: The Future of Development
Cursor 3.0’s evolution signals a significant trend in software development: the move from direct code manipulation to orchestrating AI systems. By abstracting complex coding tasks to AI agents, developers can focus more on high-level architecture, design, and problem-solving. This could dramatically speed up project timelines, enable smaller teams to tackle larger projects, and potentially lower the barrier to entry for creating sophisticated software.
The introduction of powerful, albeit initially opaque, AI models like Composer 2, combined with an interface designed for agent management, suggests a future where AI plays an even more central role. While the transition may be jarring for some developers accustomed to traditional workflows, the potential gains in efficiency and productivity are substantial. The ability to manage multiple AI agents across various environments from a single platform represents a new frontier in how software is built.
Sponsorship Message: Blacksmith for Faster CI/CD
This video was sponsored by Blacksmith, a service designed to accelerate GitHub Actions. Blacksmith offers a drop-in replacement for standard GitHub runners, promising to execute actions twice as fast while costing up to 75% less. They achieve this by utilizing bare metal gaming CPUs known for their high single-core performance. Blacksmith also enhances the observability of GitHub Actions, making it easier to diagnose issues in CI/CD pipelines, a crucial feature when AI agents are generating large amounts of code with minimal oversight. Interested users can try Blacksmith for free, receiving 3,000 free minutes per month upon signup.
Source: Cursor ditches VS Code, but not everyone is happy… (YouTube)
