Cursor Cloud Agents With Computer Use: AI Builds Software, Tests It, and Sends Video Proof

What happens when your AI coding assistant can not only write code, but run it, test it, record a video of its work, and hand you a merge-ready pull request? That's not a hypothetical anymore. That's a new agent-computer-use feature that Cursor has revealed.

Software development has changed a lot over the years, moving from manual compilation to IDEs, from waterfall to agile, from manual QA to automated testing pipelines, completely changing how developers work and what they can realistically build. Cursor, the AI-powered code editor developed by Anysphere, Inc., has announced a new feature that could again redefine how developers work.

In a blog post, the company announced that its cloud-based AI agents can now control their own computers, meaning AI agents can build software, interact with it, validate that it works, and produce video recordings and screenshots as proof, all without a human in the loop. It is an autonomous agent that operates an entire development environment, runs the software it builds, and iterates until it is confident in the result.

The Core Idea: Agents That Can Use What They Build

The fundamental limitation that Cursor identified with earlier AI agents was environmental. As the company put it, agents are only as capable as the environment they run in, and without the ability to use the software they have created, they hit a ceiling. Over the past several months, Cursor addressed this by giving agents their own virtual machines (VMs) with full development environments, along with the ability to test their changes and produce artifacts like videos, screenshots, and logs, so developers can quickly validate the work.

Once assigned a task, the new version of Cursor cloud agents can onboard themselves onto a codebase, execute the work, and produce merge-ready pull requests accompanied by artifacts that show what was done. Developers can also take control of the agent's remote desktop to interact with the modified software themselves without needing to check out the branch locally.

Why Use Cloud Agents Over Local Agents

Before cloud agents, Cursor's AI ran locally on the same machine as the developer. This created inherent conflicts where local agents compete with each other and with the developer for the computer's own resources. They cannot run in true parallel, and they cannot safely interact with the application they are building in an isolated environment.

Cloud agents solve this by giving each agent its own isolated VM. Multiple agents can run simultaneously without interfering with each other or with the developer's own workflow. Crucially, these agents can also build and interact with software directly in their sandbox, then iterate until they have validated their output rather than handing off the first attempt and leaving verification to the developer. Demonstrated capabilities in the company's earlier research include navigating web pages in a browser, using tools like spreadsheets, interpreting data, making decisions, and resolving issues in complex UI environments.

How Cursor Uses Its Own Agents: Four Real Examples

According to Cursor, by running cloud agents, instead of breaking tasks into small chunks and micro-managing agents, they now delegate more ambitious tasks and let them run independently.

The result: More than 30% of the pull requests merged at Cursor are now created by agents operating autonomously in cloud sandboxes.

Here is how cloud agents are being used in practice:

• Building new features: Cursor used cloud agents to help them build plugins for the Cursor Marketplace.

Example: In one example, an agent was tasked with adding source-code links to each component on a plugin's page, constructing GitHub URLs for skills, commands, rules, subagents, hooks, and MCP configurations. It implemented the feature, verified each link by navigating the imported Prisma plugin page, temporarily bypassed a feature flag for local testing, then reverted the change, rebased onto main, resolved merge conflicts, and squashed everything into a single commit.

• Reproducing security vulnerabilities: When the team needed to investigate a clipboard exfiltration vulnerability, they triggered a cloud agent directly from Slack with a prompt asking it to triage and explain the issue in detail. The agent built a working exploit, hosted it on a local backend server, executed the full attack flow inside its own VM, and reported back in the original Slack thread with a video artifact and summary of its findings.
• Handling quick fixes: For a targeted UI change like replacing a static "Read lints" label with a dynamic one that reflects actual lint diagnostics, a cloud agent implemented the fix and tested it against two cases inside the Cursor desktop app:
  • A file with type errors.
  • A clean file.

It recorded itself verifying both cases and confirmed the output matched the expected behavior before submitting.

• Testing UI comprehensively: A cloud agent was tasked with verifying that everything works correctly on Cursor's documentation site, spending 45 minutes conducting a full walkthrough. It tested the sidebar, top navigation, search, copy page button, share feedback dialog, table of contents, and theme switching, then delivered a structured summary of every feature it covered.

In Conclusion:

Autonomous AI agents are bringing big changes to how developers work. First, we had GitHub Agentic Workflow, which could help developers automate their repositories. Now, we have Cursor cloud agents with computer use that allow AI agents to not only build software but also interact with it, validate that it works, and produce video recordings and screenshots as proof, all without a human in the loop. With these consistent and actually useful improvements, AI agents are becoming involved in everyday tasks, helping developers and business professionals to automate workflows.