CursorAgent loopLevel 3: Workflows

Which Run Mode Should You Use in Cursor?

Run Modes decide when Agent can run shell commands, MCP tools and fetches without asking you. Go to Settings > Agents > Approvals & Execution to choose. Cursor's docs recommend Auto-review for most people, Allowlist when you want predictable behavior, and Run Everything only if you accept the risk.

Prompt to copy
Add this to my Auto-review settings by editing .cursor/permissions.json:

I want every [AWS CLI / kubectl / database migration] command to go through approval first.

The file format is:
{
  "autoRun": {
    "allow_instructions": [],
    "block_instructions": ["[plain-English sentence]"]
  }
}

Show me the result, and explain which of my rules would and wouldn't catch a command like [example command].
  1. 1Auto-reviewallowlist, then sandbox, then classifier
  2. 2Allowlistyour list runs, the rest asks
  3. 3Run Everythingno prompts, no sandbox
Cursor's three Run Modes. Its docs call Auto-review the safest useful setup for most people.

Try it now

  1. 1Open Settings > **Agents > Approvals & Execution** and check which mode you're on.
  2. 2Choose Auto-review. If it's grayed out on a team, an admin may have blocked its fallback model, which the docs say is Claude 4.5 Haiku.
  3. 3Ask Agent to run your tests. It worked if the command runs without a prompt and the chat shows when anything runs outside the sandbox.

Use it when

Set this once when you start using Agent seriously, and review it when you start running long or parallel agents, because approval prompts are what stall them. Tighten it for repositories with production credentials nearby.

Skip it when

Don't choose Run Everything on a machine holding production keys or customer data. Don't treat Auto-review as your protection either. Cursor's docs say plainly that it isn't a security boundary, and that the classifier can allow a call you would have blocked.

Why it works

Auto-review checks each call in order. Allowlisted calls run immediately. Other shell commands run in a sandbox when they can, which limits files and network. Anything that can't use the sandbox goes to a classifier on Cursor's backend, which can read related files to judge the call. If it blocks a call, the agent tries another approach or asks you. By default, sandboxed commands can read and write in the workspace, with network blocked until your settings open it, and paths like .git/hooks and .cursorignore are protected.

Example

Say you work in a repo that uses the AWS CLI for deployments. You pick Auto-review, then tell Agent: "I want every AWS CLI command to go through approval first." It writes a plain-English entry under block_instructions in .cursor/permissions.json, and you commit the file so the team shares it. Tests and builds keep running without interruption, and a deploy command now stops to ask you.

Going further

One setting people miss: Read Access, from Cursor 3.23, in the same section. Its default, System, lets the agent read files outside your workspace without approval, which can include other projects and dotfiles. Switch to Workspace to require approval, and add exceptions to the Read Allowlist. permissions.json and sandbox.json do different jobs: the first steers what Auto-review runs, and the second sets what a sandboxed command can reach, like network domains and extra paths. When a team defines an Auto-review configuration in the dashboard, Cursor ignores your local files.

Common mistakes

  • Using Run Everything because prompts were annoying, rather than widening the sandbox or allowlist.
  • Assuming .cursorignore protects secrets from shell commands. Cursor's docs say terminal and MCP tools can't be blocked by it.
  • Leaving Read Access on System without knowing it exists.
  • Editing local permissions.json on a team with a dashboard policy, and wondering why nothing changed.

Sources

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

Related

Beyond the course