In a major move to solidify its lead in the “vibe coding” era, Anysphere has officially launched Cursor 3. This release marks the transition of Cursor from a smart code editor into a unified workspace for agent orchestration, designed to handle the complexity of modern, high-velocity software engineering.
The launch comes at a critical time as Cursor faces intense competition from Claude Code (which recently captured 54% of the AI coding market) and OpenAI’s Codex 5.3.
1. The Core Innovation: “Agents Window”
Cursor 3 moves beyond the traditional sidebar chat. The new Agents Window acts as a command center for autonomous workflows, allowing developers to manage multiple tasks across different environments simultaneously.
- Parallel Execution: You can now run several agents at once—one refactoring a service layer, another writing unit tests, and a third managing a cloud deployment.
- Hand-off System: Seamlessly move agent sessions between your local machine, remote SSH servers, and the Cursor Cloud.
- Redesigned Interface: A “tabbed” agent view that lets you toggle between different active tasks or view them alongside your code in a split-screen “Design Mode.”
2. New Power Features: “/worktree” & “Best-of-N”
Cursor 3 introduces several native commands that change how developers interact with their version control and AI models.
| Feature | Functionality |
/worktree | Automatically spins up a git worktree for an agent to work in, ensuring your main branch remains untouched while the AI experiments in an isolated environment. |
Best-of-N | A new comparison tool that runs the same prompt through multiple models (e.g., Composer 2, Claude 4.5, and GPT-5.2) and presents the best output for you to choose. |
| Design Mode | A visual-first environment where designers and QA engineers can verify agent changes in a built-in browser without reading the underlying code. |
3. The “Composer 2” Engine
Under the hood, Cursor 3 is powered by Composer 2 (released mid-March). This proprietary model is trained exclusively on coding data, making it more token-efficient and faster than general-purpose LLMs.
- Long-Horizon Tasks: Optimized for tasks that require an agent to stay active for hours, such as migrating a legacy front-end from Vue to React.
- Real-Time RL: Uses reinforcement learning from real developer interactions rather than static benchmarks, allowing it to adapt to “vibe coding” patterns more naturally.
- Large Codebase Retrieval: Improved context windows specifically for massive, interconnected production systems with over 1 million lines of code.
4. Enterprise & Security Upgrades
With over 50,000 businesses (including Stripe, Figma, and NVIDIA) now using the platform, Cursor 3 adds critical “Day 2” operations features:
- Self-Hosted Cloud Agents: Companies can now run Cursor’s autonomous agents within their own VPC (AWS/GCP/Azure) to ensure code never leaves their secure perimeter.
- Admin Controls: New security dashboards for tracking AI-generated code attribution and managing MCP (Model Context Protocol) plugin permissions.
- MCP Integration: Full support for the Model Context Protocol, allowing agents to use third-party tools like Slack, Jira, and Notion to report progress or pull requirements.
5. Market Position & Valuation
The launch of Cursor 3 follows a massive $2.3 billion Series D round in late 2025, which valued Anysphere at $29.3 billion.
“We are building toward a future of self-driving codebases,” said co-founder Michael Truell. “Cursor 3 is the interface for that future, where the developer’s role shifts from writing every line of code to managing a fleet of specialized agents that do the heavy lifting.”
