<h2>What’s new</h2>
<p>Cursor has released Cursor 3, a significant overhaul that rethinks the interface, expands cloud agent capabilities, and adds self-hosted deployment for enterprise teams.</p>
<p>The headline change is that agents can now test their own changes, record their work via video and logs, and run in parallel across cloud-based virtual machines. Developers can run 10–20 agent tasks simultaneously. About 35% of Cursor’s pull requests are already generated by these autonomous agents.</p>
<h2>A rebuilt interface</h2>
<p>The new UI was built from scratch starting in December, after the team concluded that the existing IDE-centric design was increasingly awkward for agent-heavy workflows. The tension was real: even if agents write 98% of the code, the remaining 2% - viewing files, debugging, small refactors, go-to-definition, LSP support - still matters enough that you can’t just strip the editor out.</p>
<p>The solution was a separate agent window that runs alongside the IDE rather than trying to absorb it. Agent chats are now normal tabs, which makes splits, panes, and keyboard shortcuts work the way you’d expect. The team also unified the local and cloud code paths, which had diverged into an ugly <code>if (local) { ... } else if (cloud) { ... }</code> situation as cloud agent usage grew.</p>
<p>On the performance side, the new interface uses the React Compiler and has gone through more profiling and memory leak fixes than previous versions. The result is noticeably faster and less janky.</p>
<h2>Self-hosted agents for enterprises</h2>
<p>Cursor is now making self-hosted cloud agents generally available. Each agent runs in an isolated VM with a terminal, browser, and full desktop - cloning repos, writing and testing code, and pushing changes without leaving the company’s network.</p>
<p>The setup connects outbound via HTTPS to Cursor’s cloud for orchestration, model access, and UX, but no inbound ports or VPN are required. Code, secrets, and build artifacts stay inside the company’s environment. For larger deployments, there’s a Helm chart and Kubernetes operator to manage worker pools and scaling. The feature supports up to 10 workers per user and 50 per team, with larger limits available on request.</p>
<p>Early enterprise adopters include Notion and Brex. The self-hosted model is a direct response to security and compliance requirements in regulated industries, where sending source code to an external cloud is often a non-starter.</p>
<h2>The competitive backdrop</h2>
<p>The release comes as the agentic coding space gets crowded fast. Anthropic’s Claude Code has reached $2.5B in run-rate revenue, OpenAI’s Codex passed 1.5M weekly active users, and GitHub Copilot claims 26M users. Cursor, valued at $29.3B with over $1B in annualized revenue, is betting that a purpose-built interface and enterprise-grade deployment options give it an edge beyond raw model quality.</p>
<p>Cursor also recently released Composer 2, its own model built on Moonshot AI’s Kimi K2.5, though the model’s origins drew some scrutiny after not being disclosed at launch.</p>


Collections

Cursor 3 is a major release featuring a rebuilt interface designed for agent-heavy workflows, parallel cloud agents that can test their own changes and run on isolated VMs, and self-hosted deployment for enterprise teams. Agents can run 10–20 tasks simultaneously, and about 35% of Cursor's pull requests are already generated autonomously. The new UI separates the agent window from the IDE, unifies local and cloud code paths, and uses the React Compiler for better performance. Enterprise self-hosting connects outbound via HTTPS for orchestration while keeping code and secrets inside the company network, with Helm and Kubernetes support for scaling. Early adopters include Notion and Brex. The release comes amid intense competition from Claude Code, OpenAI Codex, and GitHub Copilot.

Cursor launches self-hosted cloud agents and parallel task execution for enterprise teams