Now loading...
Cursor, the AI-powered code editor, is taking a significant step forward in autonomous software development by rolling out an upgraded version of its cloud agents. These digital assistants now operate within isolated virtual machines equipped with complete development setups, enabling them to test modifications, generate demonstrations through videos, screenshots, and logs, and deliver results that developers can easily verify.
The new cloud agents are accessible across multiple platforms, including web browsers, mobile devices, the desktop application, Slack channels, and GitHub repositories. Once activated, they integrate seamlessly with a project’s codebase, crafting pull requests ready for merging along with visual proof of their work. Developers can even take control of the agent’s remote desktop to interact with the updated software and apply personal tweaks, all without needing to download the code branch to their own machine.
This advancement marks a pivotal change in software creation workflows, comparable to the transition from simple tab completion to real-time collaboration with AI. At Cursor, over 30 percent of merged pull requests now originate from these agents working independently in secure cloud environments.
Unlike local agents, which often clash over resources or generate incomplete code on the first try, cloud versions thrive in their dedicated virtual spaces. This setup allows multiple agents to run simultaneously without interference, and they can build, execute, and refine software right in their sandbox until the output meets quality standards. Early experiments, as shown in demonstration videos, highlight agents browsing websites, handling spreadsheet tools, analyzing information to make choices, and troubleshooting intricate user interfaces.
Internally at Cursor, teams have relied on these cloud agents for the past month, transforming their approach from micromanaging small tasks to assigning larger, more complex projects. For instance, in developing plugins recently introduced on the Cursor Marketplace, an agent was tasked with adding source code links to each plugin component’s display page. It tracked file locations like markdown documents for skills or JSON files for hooks, constructed GitHub URLs, integrated the links into the frontend with an appropriate icon, and tested everything against a sample repository.
The agent captured footage of itself verifying the links by clicking through components on an imported plugin page. To ensure thorough local validation, it temporarily disabled a feature flag blocking access to the marketplace, then restored it before finalizing the code. It also handled rebasing onto the main branch, fixed conflicts, and consolidated changes into one commit.
In another application, a cloud agent summoned via Slack dissected a clipboard data theft vulnerability in detail. It constructed an HTML exploit page hosted on a local server, simulated the attack in Cursor’s built-in browser by copying a test identifier to the clipboard and triggering the exfiltration, and documented the process with screenshots and a video walkthrough. The result was committed as a demo file, providing a clear illustration of the security issue.
For routine improvements, agents have tackled updates like swapping a fixed “Read lints” label for a dynamic version that reflects actual error counts, such as “No linter errors” or “Found N errors,” complete with matching styles. Testing involved running scenarios in the desktop app with error-prone and clean files, with recordings confirming the label’s correct behavior.
Even user interface validation benefits from this technology. One agent conducted a comprehensive 45-minute exploration of the documentation site at cursor.com/docs, checking elements like the sidebar, navigation, search functionality, copy buttons, feedback forms, contents tables, and theme toggles, then summarizing its findings.
As agents shoulder more of the coding burden, developers at Cursor report shifting their focus toward guiding strategy and approving releases. Looking ahead, the company envisions self-sustaining codebases where agents not only create but also integrate changes, oversee deployments, and oversee live systems. Achieving this will involve refining tools, enhancing AI models, and improving how agents collaborate, with immediate efforts centered on multi-agent orchestration and systems that improve through accumulated experience.
Users can begin experimenting via cursor.com/onboard, where an agent sets up and records a sample demo, or dive deeper into the documentation.
