Cursor Cloud Agents: Build and Test in Isolated VMs

What Are Cursor Cloud Agents

Cursor Cloud Agents are autonomous coding agents that execute tasks inside cloud-hosted virtual machines. Unlike local Cursor, which runs the AI model alongside your editor on your machine, Cloud Agents offload the entire execution to a remote VM. You describe a task in natural language, the agent spins up a dedicated environment, clones your repository, installs dependencies, and begins working. When finished, it creates a pull request targeting your specified branch.

The key distinction from local AI coding assistants is that Cloud Agents run asynchronously. You do not need to keep your editor open, watch the agent work, or stay online. Assign a task, continue with your own work, and receive a notification when the PR is ready. This makes Cloud Agents particularly effective for parallelizing independent tasks across a team or backlog.

VM Isolation Architecture

Each Cloud Agent task receives a fresh virtual machine with its own filesystem, terminal sessions, network stack, and package manager environment. The VM clones your repository at the current HEAD of your specified branch, runs your setup commands (such as npm install or pip install -r requirements.txt), and then begins executing the assigned task.

This isolation model mirrors how CI/CD pipelines work: each run gets a clean environment, executes in isolation, and produces artifacts without polluting shared state. The difference is that Cloud Agents are not just running tests. They are writing the code, running the tests on that code, and iterating until the tests pass.

Video Output and PR Workflow

One of the most distinctive features of Cloud Agents is the video recording attached to every completed task. The agent records its entire session inside the VM: opening files, writing code, running terminal commands, reading error messages, and iterating on solutions. This recording is attached to the pull request as a playable video.

The video output addresses a common concern with autonomous coding agents: the black box problem. When an agent produces a diff, you see what changed but not why. The video provides the reasoning trail. You can watch the agent encounter an error, read the stack trace, decide on a fix, implement it, and verify the solution. This transparency makes code review faster because you understand the intent behind each change.

Performance: 30%+ PR Merge Rate

Cursor reports that Cloud Agents achieve a PR merge rate above 30% on internal benchmarks. This means roughly one in three pull requests generated by Cloud Agents passes CI, clears code review, and merges to the target branch without requiring significant manual intervention. For context, this is not a pass rate on toy benchmarks like SWE-bench. It measures real production codebases where the PR must satisfy linting, type checking, test suites, and human review.

A 30% merge rate might seem low in isolation, but the economics shift when you consider that Cloud Agents work asynchronously and in parallel. Assigning five tasks simultaneously means an expected return of 1-2 merged PRs with no hands-on development time. For teams with large backlogs of well-defined tickets, this translates directly into throughput gains.

Supported Languages and Frameworks

Cloud Agents run inside Linux-based VMs with standard development toolchains pre-installed. Any language and framework that builds in a typical CI environment will work. The agent reads your project configuration files, installs dependencies from lock files, and adapts to your toolchain automatically.

The agent performs best on strongly-typed codebases with good test coverage. TypeScript projects with Vitest or Jest suites, Python projects with Pytest, and Go projects with built-in testing all provide the feedback loops that let the agent self-correct during execution. Dynamically-typed codebases without tests still work, but the merge rate is lower because the agent has fewer signals to validate its own output.

Cloud Agents vs Local Cursor

Cloud Agents and local Cursor are complementary tools, not replacements for each other. Understanding when to use each one determines how much value you extract from both. The decision comes down to task type, interaction model, and isolation requirements.

The practical workflow for most teams combines both. Use Cloud Agents to parallelize well-defined tickets from your backlog: bug fixes, test additions, documentation updates, dependency upgrades, and small feature implementations. Use local Cursor for complex architectural decisions, multi-file refactors that require ongoing dialogue, and any task where you need to guide the AI step by step. For comparison with terminal-based agents, see our Claude Code Remote Control guide, which covers a different execution model for autonomous coding.

Pricing and Access

Cloud Agents are available on Cursor Pro and Business plans. The pricing model ties to compute credits consumed by VM runtime rather than a flat per-task fee. A typical task that runs for 10-15 minutes in the VM consumes a moderate number of credits. Longer tasks (such as large refactors or complex feature implementations) consume proportionally more.

For teams evaluating whether Cloud Agents justify the cost, the calculation is straightforward: compare the credit cost per merged PR against the hourly cost of developer time for equivalent work. If a Cloud Agent produces a merged PR for the equivalent of $5-15 in credits, and the same task would take a developer 1-3 hours at $75-200 per hour, the ROI is significant even at a 30% merge rate. For more on building and selling AI agent solutions, see our guide to building and selling custom AI agents.

Putting It All Together

Cursor Cloud Agents represent a meaningful shift in how development teams can approach their backlog. The combination of VM isolation, asynchronous execution, video-recorded sessions, and automatic PR creation turns well-defined tickets into a parallelizable queue. You assign tasks, the agents execute in isolated environments, and you review the results when they arrive.

The 30%+ merge rate is an honest benchmark: not every task will produce a mergeable result, but the cost of a failed attempt is limited to compute credits and review time. For teams with large backlogs of well-scoped tickets, the math works in your favor. For complex architectural work, local Cursor and tools like Claude Code with Claude Opus 4.6 or GPT-5.2 remain the better choice because they support the real-time dialogue that nuanced decisions require.

The practical path forward is not choosing one tool over another. It is building a workflow that assigns the right tasks to the right execution model: Cloud Agents for parallelizable, well-defined work, and local AI tools for interactive, complex sessions.

Frequently Asked Questions