Cursor: Cloud Agents with Computer Use

Cloud Agents Now Control Their Own Computers

Cursor introduced a major evolution to its cloud agents on February 24, 2026: agents can now use the software they create to test changes and demonstrate their work before handing anything off to a developer. Rather than completing a task and immediately requesting review, agents iterate inside their own environment until they have validated output.

Isolated Virtual Machines for Every Agent

After onboarding onto a codebase, each cloud agent runs inside its own isolated virtual machine equipped with a complete development environment. This architecture eliminates the resource contention that local agents face — multiple agents no longer compete for CPU, memory, or disk on a developer's laptop. Instead, cloud agents operate in parallel, each with its own dedicated compute environment, enabling 10 to 20 concurrent tasks where previously one to three was the practical limit.

Artifacts That Prove the Work

Every cloud agent session produces a set of artifacts alongside the pull request:

• Video recordings of the agent implementing and testing changes in real time
• Screenshots capturing successful states and UI outcomes
• Process logs documenting each step taken

These artifacts let developers quickly audit exactly what an agent did and why, without re-running the work themselves. Pull requests arrive merge-ready, complete with supporting evidence.

Real-World Capabilities Demonstrated Internally

Cursor's own engineering team tested cloud agents on a range of authentic tasks:

Plugin development. Agents built source code linking functionality for the Cursor Marketplace, then verified that GitHub links worked correctly by navigating the plugin interface themselves.

Security research. An agent triaged a clipboard exfiltration vulnerability reported via Slack, constructed a proof-of-concept exploit page, hosted it locally, and recorded the full attack flow as evidence.

UI quality assurance. Agents conducted 45-minute walkthroughs of documentation sites, systematically testing navigation, search, and theme switching — logging findings in structured reports.

Runtime debugging. Agents implemented dynamic lint result labels in the desktop application, testing multiple error state scenarios to confirm correct display behavior.

Availability and Access

Cloud agents with computer use are available across every Cursor surface: the web interface, desktop application, mobile apps, Slack integration, and GitHub. Developers can get started at cursor.com/onboard to watch an agent configure itself and produce a demo recording from scratch.

Internal Impact and the Road Ahead

More than 30% of pull requests merged internally at Cursor are now created by autonomous cloud agents. The Cursor team describes the shift in developer role as moving from implementation to direction-setting: deciding what ships rather than building it line by line.

Near-term priorities include coordinating work across multiple agents simultaneously and training models that improve from accumulated experience across past runs. The longer-term goal is what Cursor calls "self-driving codebases" — agents that can merge pull requests, manage deployment rollouts, monitor production systems, and ship features end-to-end without a human in the loop for routine tasks.