Cursor 3.0 introduced a change that many people still don’t use: Background Agents. They are agents that run remotely, in parallel, while you keep coding in your local IDE. The difference between having them active or not is the difference between “I ask the LLM to fix something and wait” and “I ask for 4 things, keep working, and after 10 minutes I have 4 reviewable PRs”.
I’ll show you how to set them up, the real requirements (yes, there’s a minimal cost) and where they’re worth it.
What they are exactly
Background Agents are remote instances of Cursor running on Anthropic/Cursor machines, not on your Mac. They can:
- Read and modify your repository (with GitHub permission).
- Run tests and commands.
- Create pull requests.
- Work for several hours on a large task without blocking your IDE.
Along with them came the Subagents in v2.4 (January 2026): parallel agents within a single task, each with its own context. Useful when a large task can be split into independent pieces.
Requirements before starting
This is what the official documentation asks for, no surprises:
- Paid Cursor account (Pro or Business).
- Privacy Mode disabled. Background agents need to transmit code to remote infrastructure.
- Usage‑based spending active with a minimum of $10‑20 balance. Each agent consumes credits.
- GitHub connected with read‑write permissions to the repo where they will work.
If your work is very sensitive and you can’t disable Privacy Mode, Background Agents aren’t for you. Use the normal Composer mode on your machine.
Tutorial · Your first Background Agent in 5 steps
-
Enable access. In Cursor:
Settings → Beta → Background Agents → Enable. Link GitHub if you haven’t already. -
Add balance.
Settings → Billing → Usage‑based pricingand add a minimum of $10. It won’t charge you for everything: only what it consumes. -
Launch the task. In any project:
Cmd+Shift+P → "New Background Agent". It asks for repo and base branch. - Write the prompt. Be specific: “Add unit tests to all methods of UserService.ts. Use the existing pattern from FooService.test.ts as reference. Don’t touch other files. Create a branch feat/user-service-tests.”
- Keep working. The agent runs remotely. You receive a notification when it finishes (15‑60 min depending on the task). Review the PR before merging.
Cases where they’re worth it
- Massive refactor. “Change all calls to deprecated_api to new_api across the monorepo.”
- Test coverage. “Add missing tests to these 6 files following the pattern of the 7th.”
- Documentation. “Add JSDoc to all public methods in this folder.”
- Routine updates. “Upgrade all dependencies in package.json to their latest minor version.”
- Isolated bugs. When you’ve already identified the cause but fixing it by hand would take an hour across 8 files.
Cases where NOT
- Architectural decisions. The agent doesn’t know your product context. That’s up to you.
- Critical security code. Auth, cryptography, payments: review yourself.
- Poorly defined tasks. If your prompt is ambiguous, the agent works in one direction and you discover an hour later it wasn’t the right one.
Subagents: the little‑used trick
Subagents let you split a task into pieces that run in parallel. Real example: “migrate components from an old library to a new one” can become 5 subagents, one per folder. Each has a clean context and they don’t step on each other.
They are activated with the syntax: @subagent task: "..." inside the main task.
Expected real cost
A typical 30‑60 min task with Claude Sonnet behind it costs between $0.80 and $2.50. If you launch 5 agents in parallel per week, we’re talking about ~ $50/month on top of the Cursor Pro plan. Compare it with your billable hour and decide.
FAQ
Do I need to be Cursor Pro?
Yes. Background Agents are not available on the free plan. Cursor Pro ($20/month) or Business.
What happens if I forget and leave an agent running?
Cursor has hard timeouts (several hours). But the balance is consumed. Enable billing alerts.
Can I use my own model (Claude/GPT)?
The Business plan lets you bring your own API key. In Pro you use the models that Cursor includes.
Does it work on private repos?
Yes, as long as GitHub is connected with permissions to the repo.
What’s the difference with normal Composer?
Composer runs on your Mac and blocks your IDE while it works. Background Agent runs remotely and frees you.
Next step
If Background Agents sound like the next level but you’re still in the “using Cursor halfway” stage, the advanced AI course covers Cursor, Copilot and Claude Code in depth. And if you want to compare AI programming tools, check out the analysis Copilot vs Cursor vs Claude Code · 2026.
Want to offer this as a service? AI Developer Guide with real rates and paths.
Sources: DeployHQ · Cursor guide 2026 · Cursor Background Agents guide.
