Cursor gives its agents isolated computers to build and test code

Cursor's coding agents don't just write code anymore. They now have their own cloud computers to run it, click around, and prove it actually works before asking you to review.

Cursor rolled out its new cloud agents, giving each autonomous agent an isolated virtual machine equipped with a full development environment. Instead of hogging your local laptop resources, these agents can build projects, open browsers, test user interfaces, and package up ready-to-merge pull requests complete with video recordings, screenshots, and logs.

You can summon these cloud agents from almost anywhere, including the web, mobile, the desktop app, Slack, and GitHub. According to Cursor, this setup is already running deep inside their own engineering team: more than 30% of all PRs merged at Cursor are now created autonomously by these sandbox agents.

Giving AI a virtual desktop

Running agents locally has always been a pain. They compete with you for CPU cycles, fight over ports, and often break local configs when running simultaneous tasks. Moving them into cloud VMs solves that bottleneck, allowing developers to spin up multiple agents in parallel without their laptop fans spinning out of control.

More importantly, the agents have full computer use capabilities inside their sandbox. If an agent builds a new UI feature, it boots up the local dev server, opens a browser, and clicks through the buttons to verify the flow. It even knows how to temporarily toggle feature flags for testing and clean up its git history before pushing.

Cursor shared several internal examples. In one case, an engineer asked an agent inside Slack to reproduce a clipboard security vulnerability. The agent spun up a backend, crafted an exploit page, executed the attack in the browser, and dropped a video artifact back into the Slack thread showing the stolen token. In another test, an agent spent 45 minutes exhaustively clicking through Cursor's entire documentation site to test navigation, dark mode, and search.

If something looks slightly off, developers do not need to pull the branch locally. You can simply open the remote desktop session inside Cursor and take control of the agent's VM directly.

Why it matters

We think this is the single biggest step toward making asynchronous software development actually practical. The main friction with AI coding has never been typing speed; it has been the verification tax. Having to pull a branch, install dependencies, and test whether the AI hallucinated a button is exhausting. By forcing the agent to record a video proof of its work inside a sandbox, Cursor removes that friction.

It also fundamentally shifts the competitive landscape for AI developer tools. The competition is no longer just about who has the best autocomplete model. It is now an infrastructure game around sandbox orchestration, browser automation, and visual verification artifacts. Cursor is pushing rivals like GitHub Copilot to treat AI not as a text completion engine, but as an autonomous junior engineer operating a remote workstation.

The real question mark is economics. Running persistent cloud VMs with full graphical environments for hours at a time is expensive. If Cursor can keep costs manageable while ensuring strict security boundaries around agent execution, this could easily become the default way software gets built.