Cursor's Cloud Agents Can Now Control Their Own Computers
Cursor's cloud agents gain the ability to interact with browsers and desktop applications inside isolated VMs, enabling autonomous testing, demos, and merge-ready pull requests.
Cursor announced on February 24, 2026 that its cloud agents can now operate their own virtual machines โ complete with browsers, desktop applications, and full development environments. The result: agents that donโt just write code, but build it, test it, and record video proof of their work before handing you a merge-ready pull request.
Beyond Code Generation
The core limitation of coding agents has always been the gap between generating code and verifying that it actually works. Cursorโs cloud agents close that gap by running inside isolated VMs where they can spin up dev servers, navigate web pages, interact with UI elements, and capture screenshots and videos as artifacts.
According to Cursor, more than 30% of merged PRs at the company are now produced by agents operating autonomously in cloud sandboxes โ a figure that suggests this is no longer a proof-of-concept.
Available Everywhere
Cloud agents can be triggered from the Cursor web app, desktop client, mobile, Slack, and GitHub. Each agent onboards itself onto your codebase automatically and runs in its own isolated environment, meaning you can launch multiple agents in parallel without them competing for local resources.
Developers can also connect to an agentโs remote desktop to interact with the modified software directly, or make edits without checking out the branch locally.
How Cursor Uses It Internally
Cursor shared several internal use cases that illustrate the capability:
Feature development: The team used cloud agents to build parts of the recently launched Cursor Marketplace plugin system. An agent implemented GitHub source links for plugin components, then verified the links by navigating to the Prisma plugin page and clicking through each component. It handled feature flag management, rebasing onto main, resolving merge conflicts, and squashing commits autonomously.
Vulnerability reproduction: A cloud agent was launched from Slack to triage a clipboard exfiltration vulnerability. The agent built an HTML exploit page, hosted it on a local server, loaded it in Cursorโs in-app browser, and recorded the complete attack flow on video.
Getting Started
To try cloud agents with computer use, visit cursor.com/onboard. The agent will configure itself against your codebase and record a demo of its capabilities.
The addition of computer use to cloud agents represents a shift from โwrite code and hope it worksโ to โwrite code, verify it works, and prove it.โ For teams already using Cursorโs agent features, the upgrade path is straightforward โ and for everyone else, the 30% PR stat might be reason enough to pay attention.
Related Articles
Cursor Adds Subagents, Skills, and Image Generation - Major Agent Enhancements
Cursor implements subagents, agent skills, and image generation. Parallel execution, improved context management, and dynamic skill application enable handling of more complex, long-running tasks.
Cursor Launches Long-Running Agents in Research Preview - Autonomous Execution for Complex Tasks
Cursor introduces long-running agents capable of autonomous planning and execution without human intervention. Available for Ultra, Teams, and Enterprise plans.
Cursor 2.5 Brings Plugin Marketplace, Sandbox Network Controls, and Async Subagents
Cursor releases version 2.5 with a plugin marketplace for extensibility, granular network access controls for sandboxed environments, and asynchronous subagent execution for parallel processing.
Popular Articles
Cursor Automations Release - Building Always-On Agents with Triggers and Schedules
Cursor releases Automations, a new feature for building always-on agents that run based on triggers and schedules you define.
Cursor v2.7 Release - Cursor Now Available in JetBrains IDEs
Cursor v2.7 introduces Cursor support in IntelliJ IDEA, PyCharm, WebStorm, and other JetBrains IDEs through the Agent Client Protocol (ACP).
Cursor v2.6 Release - MCP Apps and Team Plugin Management
Cursor v2.6 introduces interactive UIs in agent chats, team-based private plugin sharing, and improvements to debug mode capabilities.
Latest Articles
Claude Code v2.1.72 Released - Auto-Memory, /loop Command, and Cron Scheduling
Anthropic releases Claude Code v2.1.72 with auto-memory for context saving, /copy command, /loop command, Cron scheduling tools, Voice keybinding customization, and more.
Gemini 3.1 Pro โ Competing for Top Spot in Coding & Reasoning Performance
Google's Gemini 3.1 Pro, released in February 2026, achieves 80.6% on SWE-Bench Verified and 2887 Elo on LiveCodeBench, positioning it as a top-tier reasoning model alongside Claude Opus 4.6.
OpenAI Symphony โ From Supervising Agents to Managing Work
OpenAI releases Symphony, an open-source automation tool that automatically executes agents from GitHub Issues (Linear) and creates PRs.