Introducing Cursor 3: A Revolutionary Shift in AI Coding Tools

Anysphere has recently unveiled Cursor 3, a groundbreaking redesign of its AI coding tool. This new version significantly alters the primary interaction model, transitioning from traditional file editing to managing parallel coding agents. Unlike past iterations that were built on a modified version of VS Code, Cursor 3 features an interface developed from the ground up, offering users the flexibility to revert to a full IDE experience at any time.

The Vision Behind Cursor 3

The core idea driving Cursor 3 is rooted in the observation that developers predominantly spend their time orchestrating multiple agents instead of directly editing code. Co-founders Michael Truell and Sualeh Asif coined the term “a unified workspace for building software with agents.” This vision aligns with Truell’s concept of the third era of software development—a future where fleets of autonomous agents handle changes seamlessly.

Internal usage data from Cursor corroborates this shift. As of March 2025, users engaged in tab completion 2.5 times more than they did with agents. Now, that ratio has entirely flipped; users are now twice as likely to be running autonomous agents. Notably, 35% of merged pull requests from Cursor’s own engineering team are generated by these cloud-based agents.

Streamlined Interface for Enhanced Agent Management

One of the standout features of Cursor 3 is its sidebar interface, which displays all active agents—both local and cloud-based—in one unified panel. Regardless of whether agents are launched from mobile, web, desktop, Slack, GitHub, or Linear, users can effortlessly monitor their status and performance.

The ability to run multiple agents in parallel across different repositories is another significant improvement over the previous interface. This new capability paves the way for a more dynamic and efficient coding workflow.

Sophisticated Local-to-Cloud Handoffs

Cursor 3 also introduces an efficient local-to-cloud handoff feature. Developers can transfer an agent session running locally to the cloud, allowing work to continue even when offline or when switching to a different task. Similarly, cloud sessions can be reverted back to local for hands-on editing and testing—ensuring flexibility in workflow management. While Cursor claims the handoff process is swift, specific latency figures have not been published. The cloud execution utilizes Composer 2, Cursor’s proprietary coding model, which purportedly features higher usage limits compared to third-party models.

Expanding Functionality with a New Plugin Marketplace

To enhance the functionality of their agents, Anysphere has rolled out a new plugin marketplace, allowing teams to integrate various MCPs, skills, and subagents. For organizations seeking better governance and control, private team marketplaces for internal plugins can be established—similar to what AWS is doing with its Agent Registry.

Mixed Community Reactions: Embracing Change or Clinging to Tradition?

Community feedback on platforms like Reddit and Hacker News has been varied, sparking passionate discussions around Cursor’s shift from an IDE-first approach. Some users express concern that this new direction might lead to a disconnection from the traditional coding process.

This view makes you lose any connection to your code… I specifically stay with Cursor because it’s so good at being an IDE.

Others have articulated how the agent-first model may complicate workflows marked by frequent context switching and the mental load of interacting with multiple agents:

Reviewing and testing code, constantly switching contexts, juggling model contexts… is so mentally taxing that it’s practically impossible to achieve any sort of flow state.

This tension in product design—between agent-first and code-first approaches—has led some to conclude that trying to accommodate both perspectives might ultimately frustrate different segments of the user base:

Agent-first needs ambient, background autonomy. Code-first needs precise, synchronous control.

Potential Concerns: Vendor Lock-In and Costs

Concerns about vendor lock-in have also surfaced in the community. Users have expressed a desire for an ideal command center to manage all AI agents without being tied to a single vendor. In response, Cursor moderators have confirmed that their platform supports models from various vendors.

The conversation surrounding costs also remains salient. Users transitioning from Cursor to competitors like Claude Code have reported significant savings, and some estimate that different harnesses can lead to dramatically varied token budgets for similar workflows:

The proper agent command center I would want to use is the one that I could manage all AI agents… not lock into one vendor.

Cursor 3 vs. Competitors

Certainly, this new release positions Cursor 3 more directly against alternatives like Claude Code and GitHub Copilot’s agent mode. The approach varies between these tools: Claude Code operates as a CLI utility from the terminal; GitHub Copilot stays embedded in the IDE; whereas Cursor has chosen to construct a dedicated interface that relegates the IDE to a fallback option.

Getting Started with Cursor 3

Cursor 3 is now available to users. To access the new interface, users simply need to upgrade their existing Cursor setup and run the command sequence Cmd+Shift+P to open the Agents Window.