Approvals, privacy and safety
Why your agent asks for approval, what's safe to share with it, and how to keep it from oversharing in group chats.
In this tutorial
Your agent can do real things: run commands on its computer, change files, and act in the apps you connect. That's what makes it useful, and it's why it sometimes stops to ask permission. This page covers what approval requests mean, what's safe to share with your agent, and how to keep it from saying too much in group chats.
"Approval required": what it means
Nothing is broken. Before your agent runs something sensitive, it pauses and asks you to approve the exact command. Common reasons:
- Installing software or tools on its computer
- Commands that need admin-level permissions
- Deleting, moving or overwriting files
- Touching files outside its normal workspace
While it's waiting, it won't carry on with the task. If your agent seems stuck, check whether it's asked you something.
How to approve safely
- Read the whole command first. Does it match what you asked for? Is anything extra tacked on?
- Use the exact approval line your agent gives you. Copy it as it is. Don't shorten it or retype it; if the command is changed even slightly, it won't match and you'll be asked again.
- Approve once unless you're sure. A one-time approval covers that one command. That's usually all a task needs, and it keeps the guardrails on.
Say no when the command has nothing to do with your request, deletes things you didn't ask to delete, or you just don't understand it. You can always ask: "Explain what this command does in plain English before I approve it."
Why approval keeps coming back
A one-time approval covers one exact command. You'll be asked again when:
- The agent moves on to a different command (step 2 of a task needs its own approval)
- The approval line was edited when you copied it
- You've moved to a different chat, where earlier approvals don't carry over
That's the safety system working, not a loop. If a task needs many steps, ask your agent to tell you up front which commands it will need.
What's safe to share with your agent
Your agent runs on its own private computer, and your chats, files and memory stay with it. But when it thinks, your messages are sent to the AI model that powers it. So use the same judgment you'd use with any online tool.
| Safe to share | Share carefully | Don't paste raw |
|---|---|---|
| Public website copy | Internal plans and strategy | Passwords and API keys |
| General process notes | Unpublished pricing | Card or bank details |
| Made-up or anonymised examples | Customer feedback (without names) | Health or financial records |
| Your own writing and ideas | Contracts with names removed | Customers' names with contact details |
Redact first. Most tasks work just as well with placeholders: replace a client's name with [CLIENT_A] and their email with [CLIENT_EMAIL]. Share only the part of a document the task needs.
Connect apps instead of pasting passwords. To let your agent use Gmail, Notion or GitHub, connect them on the Integrations tab rather than giving it your login. See connect your agent to your apps.
Group chats: keep private things private
If your agent is in a Discord server, a Slack channel or a Telegram group, everyone there can see its replies. Without clear rules, an over-helpful agent can mention things from your private notes or memory.
Add rules like these to your agent's instructions (see give your agent a personality for where they live):
In group chats, never reveal private notes, memory, files, secrets or personal details. Give general answers only. If a request could expose personal information, say "Let's continue this in a private chat" instead.
Then test it in a group before relying on it: ask it something it shouldn't answer and check that it refuses.
If it already shared something: tell it to stop, move the conversation to a private chat, and note exactly what was exposed. If a password, token or key appeared, change it straight away in the service it belongs to.
Frequently asked questions
Why does my agent ask for approval before running commands?
It's a safety feature. Sensitive commands, like installing software or deleting files, need your OK first so the agent never does something risky without you knowing.
Is it safe to give my agent client data?
Share the minimum the task needs, and replace names and contact details with placeholders. Never paste passwords, card details or regulated records such as health information.
Do my files leave my agent's computer?
Your files are stored on your agent's own computer. When the agent works on something, the relevant text is sent to the AI model to process, which is why it's best not to include secrets.
My agent said something private in a group chat. What now?
Stop it, move to a private chat, and change any password or key that was shown. Then add a group privacy rule to its instructions and test it.
Can I turn approvals off?
Approvals protect your agent and your data, so we recommend keeping them on. If one task needs many approvals, ask your agent to list the commands it plans to run so you can approve them knowingly.