85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is a background agent?

A background agent is an AI agent that works on a task asynchronously in a remote or isolated environment, without a person watching each step, and hands back its result as a branch, pull request or report.

Checked

Most people first meet coding agents in an editor, watching each step. A background agent takes the same task and runs it elsewhere: in a cloud machine, a CI runner or a sandbox. The person starts it, moves on, and reviews the output later.

The term is mostly used for coding. Cursor shipped the feature as Background Agents, since renamed Cloud Agents. GitHub's documentation describes Copilot researching a repository, planning changes and implementing them in the background. OpenAI's harness engineering post describes recurring background Codex tasks that scan a codebase and open clean-up pull requests.

How the products run them

  • Cursor Cloud Agents

    Isolated cloud machines with the cloned repository, dependencies, secrets, start-up commands and network access. Agents run in parallel without the laptop staying online, and can be started from the editor, the web, Slack, GitHub, Linear or an API.

  • GitHub Copilot

    Work runs in a cloud development environment. Sessions on GitHub Actions have a maximum execution time of 59 minutes, and Copilot works on one branch at a time with one pull request per task.

  • Scheduled agents

    OpenAI's team runs background Codex tasks on a cadence to find code that breaks the repository's golden principles and open targeted refactoring pull requests.

What comes back

The usual output is a branch and a pull request. Cursor's documentation adds artifacts such as screenshots, videos and logs showing what changed and how it was checked, and the option to take over the agent's remote desktop. Repository rules still apply: GitHub notes that branch protection can block Copilot from creating or updating pull requests unless an admin grants a bypass.

The risks to manage

  • No one is watching

    Anything the agent is allowed to do, it may do without a person seeing it first. Sandboxes, outbound network limits and scoped secrets replace step-by-step approval.

  • Secrets in the environment

    Cursor injects secrets only when an agent starts and lets admins restrict outbound domains. Fewer and narrower secrets limit what a prompt injection could reach.

  • Review becomes the bottleneck

    OpenAI's team found human QA capacity became the constraint as agent output grew. Small, focused pull requests and automated checks keep review possible.

  • Cost

    Cursor bills cloud agents at the selected model's rate and asks for a spend limit on first use. Unattended runs need budgets.

What it is confused with

  • Long-running agent

    That term is about duration. A background agent can finish in ten minutes; what defines it is that nobody watches.

  • Subagent

    A subagent is spawned by another agent inside one run. A background agent is started by a person or a schedule.

How Motiontic's engineers apply it

Motiontic's engineers give background agents the same guardrails as a new contractor with repository access: a sandbox, a narrow set of secrets, protected branches, and a checklist the pull request must pass before a person reviews it. Recurring clean-up tasks come after the team trusts one-off ones.

An interactive coding agent compared with a background agent

An interactive coding agent compared with a background agent
Interactive agentBackground agent
Where it runsThe developer's machine or editorA cloud machine, CI runner or sandbox
Human roleWatches and steers each stepStarts the task, reviews the result
OutputEdits in the working copyA branch, pull request or report
Main controlApproval promptsSandbox, network limits, scoped secrets, branch rules
Runs in parallelRarelyOften, several at once

Common questions

What are background agents in Cursor?

Background Agents were Cursor's remote coding agents, now called Cloud Agents. They run in isolated cloud machines with the repository and its dependencies, and return branches and pull requests.

Are background agents safe?

They are as safe as their environment. Isolated machines, restricted network access, narrowly scoped secrets and protected branches matter more than for an agent a person is watching.

What is a background agent good for?

Well-specified tasks with a clear check, such as fixing a failing test, small features, dependency updates and recurring clean-up, where a person can judge the result from the pull request.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.