Approvals and safety
Approvals are how Interpreter stays fast without surprising you. Sensitive actions pause for confirmation with the exact change shown, so you can decide in a second or two without slowing the agent down.
What triggers an approval
Anything that changes the world outside a read-only inspection. By default this includes:
- writing or overwriting files outside what was explicitly asked
- sending messages, emails, or chats
- submitting forms in the browser
- deleting files or data
- making purchases or any irreversible commit
- running shell commands
- calling external APIs through configured integrations
Reading files inside the workspace, inspecting open windows, and scrolling through pages do not require approval. The line is "would I want to see this before it happens?"
Approval policy modes
You choose the policy per agent in Settings. The available modes:
- Ask (default, recommended). Every sensitive action pauses for your confirmation. Best for new workflows, unfamiliar apps, and anything involving customer data, money, or external systems.
- Auto-approve. Sensitive actions run without prompting. Use this only after you have already validated the workflow end to end on a real example, and only inside a workspace you trust. Great for batch runs you have already proven out.
- Never. Sensitive actions are blocked entirely. Use this when you want Interpreter to read and report only, with no ability to change anything.
You can also approve a specific tool for the rest of the session. That tool will stop prompting until you restart the app, while other sensitive actions still ask.
What an approval prompt shows
Every prompt is designed to be readable at a glance:
- the tool name and which agent requested it
- the exact parameters being passed
- a diff or preview of the change where one applies
- a warning callout for anything irreversible
You approve, deny, or grant for the session. If the action is unsafe, deny and tell the agent what to do instead.
Patterns that work well
Give the agent explicit stopping points in your instructions. Phrasings that hold up well across tasks:
- "Stop before sending."
- "Draft only, do not submit."
- "Prepare a copy in a new file and wait for review."
- "Show me the diff, then ask before saving."
These cooperate with the approval system rather than fighting it. The agent will plan around the pause instead of trying to push through.
The first-run rule
When you build a new workflow, run it on a single example with Ask turned on. Watch every approval. Look at what the agent actually wants to do, not what you assumed it would do. Fix the prompt, the workspace, or the scope until the prompts look exactly right.
Once the first run is clean, batch the rest. Switch to auto-approve only after the workflow has earned it.