Cursor Background Agents

What Cursor Background Agents do

Cursor Background Agents are asynchronous cloud workers for software engineering. Instead of keeping an agent tied to the local editor, a developer delegates a scoped task against a repository and lets it continue remotely. Cursor connects cloud agents with Composer, MCP servers, skills, hooks, Bugbot review, and team automations. Teams can share rules and context rather than requiring every developer to recreate the same setup.

This model works for tasks with clear acceptance tests: add tests, upgrade a dependency, apply a mechanical refactor, investigate a failing build, or prepare a pull request for a specified issue. It is weaker for ambiguous architecture decisions, incidents requiring live context, or changes where the agent cannot execute the relevant environment. Compare GitHub Copilot Agents, Devin, and local collaboration through Aider.

Pricing

The agents pages were blocked behind a JavaScript challenge during curl research, so manual verification remains enabled. Cursor’s main pricing page was readable and showed Hobby for free, with limited Agent requests and Composer. Individual Pro starts at $20 per month and includes extended Agent limits, frontier models, MCPs, skills, hooks, cloud agents, and Bugbot on usage-based billing. Teams Standard starts at $40 per user per month and adds shared context, automations, administration, analytics, privacy mode, SAML/OIDC SSO, and agentic review. Enterprise is custom-priced and adds pooled usage, SCIM, audit logs, service accounts, repository/model/MCP controls, and network, browser, and auto-run policies.

The page says each plan includes model usage and on-demand usage is billed in arrears after the included amount is consumed. Therefore, $20 or $40 is not necessarily total cost. During a pilot, record model spend per merged pull request, failed runs, reruns, and review time.

Security and practice

A cloud agent needs repository access and may execute code from the branch it changes. Use short-lived credentials, restrict network egress, prevent production-secret access, and require protected-branch review. Enterprise policies do not replace repository permissions and CI checks. Treat MCP servers as additional code and data access: allowlist them, scope credentials, and log calls.

Define “done” in every task: files in scope, commands to run, expected behavior, and forbidden changes. Require tests, static analysis, and a readable summary. Reject huge mixed-purpose diffs. Background Agents can improve throughput when work is parallelizable and reviewable; they can also create a queue of plausible but wrong pull requests. Use the AI coding agents comparison and Cursor profile to compare cost and controls.

Final evaluation check

Before committing, assign an owner, a four-week test period, a baseline, and a measurable pass threshold. Record successful completions, failures, retries, human correction time, total fees, and downstream usage costs. Recheck the vendor price and contract on the purchase date because features, allowances, and promotional terms can change after this research date.