CursorAgentsLevel 4: Multi-tool systems

How Do You Hand a Task to a Cursor Cloud Agent?

Background agents are now called Cloud Agents. Select Cloud in the dropdown under the Agent input, go to cursor.com/agents, or mention @cursor in Slack, GitHub or Linear. Each runs on its own virtual machine, so you can close your laptop and review a pull request later.

Prompt to copy
Task: [one bounded change, for example: add a `deletedAt` soft-delete field to the Order model]
Scope: [model, migration, repository, and the two queries that list orders]
Follow the pattern already used for: [`archivedAt`, in @file]
Done when: [`npm test` passes and listings exclude soft-deleted orders]

Open a PR when finished. In the PR description, say what you tested, and attach a screenshot or log showing the result. If something needed for testing is missing, say so rather than skipping the test.
  1. 1Describe the task
  2. 2Clones repo, makes branch
  3. 3Works on its own VM
  4. 4Opens a PR
  5. 5You review
How Cursor's course describes a Cloud Agent run. You review a pull request, not a live session.

Try it now

  1. 1Commit your work and push it, since a cloud run starts from the remote repository.
  2. 2Open Agent, select Cloud in the dropdown under the input, and send a small task using the prompt above.
  3. 3Close the panel and do something else. It worked if you get a pull request with the change, plus artifacts showing how it was checked.

Use it when

Cursor's course suggests the tasks you'd otherwise add to a to-do list: side-track bug fixes, test coverage for existing code, documentation updates and refactors. They should have a clear finish and a check the agent can run itself, such as a test command.

Skip it when

Work you need to steer closely, or design decisions you haven't made. Also think twice if your repo can't run on a clean machine: if it's hard for a human to test without internal services, it's hard for the agent. Admins must also connect source control first, and you need read-write access to the repo.

Why it works

A Cloud Agent clones your repo, works on a separate branch, and has a full environment where it can build and test. Since the machine is separate, its commands never touch yours, and you can run as many as you like in parallel. It attaches screenshots, videos and logs to the PR, so you can judge the work without checking out the branch. You can also take over its remote desktop to try the software yourself. On The X Subscription, Cloud Agents start at Premium; Free and Plus don't include them.

Example

Say you notice while reviewing that 30 call sites still use a deprecated fetchUser(). You don't want to stop your current task, so you comment @cursor migrate these to getUser() and run the tests on the relevant GitHub issue. It works on its own branch and opens a PR with the test log attached. You review it between other work, and comment on the PR with changes you'd like.

Going further

"Move to Cloud" doesn't snapshot local work. Cursor's help says the cloud agent starts from a clean Git state with your conversation but not your dirty files, so commit or stash first. In the CLI, prefix a message with & to hand it to the cloud. Cloud commits are authored as the agent rather than as you, which can trip review rules that check the author. On Teams and Enterprise, agents also try to fix GitHub Actions failures on their own PRs. Turn that off for one PR with the comment @cursor autofix off.

Common mistakes

  • Skipping environment setup, then getting code the agent never ran. See environment setup.
  • Moving a chat to the cloud with uncommitted changes and expecting them to come along.
  • Reading only the summary. Open the artifacts and the diff, because Cursor's course warns passing tests don't guarantee the code is right.
  • Searching settings for "background agents". The feature is now Cloud.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course