The Agent (default mode of the side pane) turns Cursor from an editor into an autonomous coding agent. It follows a reasoning + action loop:
Cursor's agent ships with a full set of tools — and there's no limit on how many tool calls it can make in a session:
| Tool | What It Does |
|---|---|
| Search | Semantic codebase search plus exact grep matching |
| Read & Edit | Read files (including images) and make precise edits |
| Shell | Run terminal commands — install deps, run tests, git operations |
| Web | Search the web and fetch pages for up-to-date docs |
| Browser | Control a browser to test and verify UI changes |
| Image generation | Generate image assets, saved into your project's assets folder |
| Ask questions | Pause and ask you for clarification when requirements are ambiguous |
What veterans knew as "YOLO Mode" has matured into auto-run with sandboxing. The agent can execute commands without prompting when they run inside Cursor's sandbox; risky or out-of-sandbox operations still require your approval, and an auto-review classifier flags dangerous shell, MCP, and fetch calls before they execute. In the CLI, toggle with /sandbox or the --sandbox enabled|disabled flag.
Need more than one agent? Open the Agents Window (Command Palette → "Open Agents Window") — a dedicated multi-agent dashboard where you run parallel agents across multiple workspaces, review diffs, manage PRs, and hand tasks off between your machine and the cloud. Git worktrees give each agent an isolated checkout so parallel work never collides.
& to hand the task straight to a Cloud Agent — it keeps running after you close the editor. More in the Cloud Agents module.Yesterday's refactor session got interrupted. Use the Cursor Agent CLI to pick a model explicitly, list previous conversations, and resume where you left off.