What should an agent ask before doing?

Before you start

• Check current account access, plan and task requirements.

Write an actionable boundary

“Ask before external email” needs a definition of external, the allowed recipients, whether attachments count and how a approval expires. Make these explicit before adding unattended execution.

Check where it is enforced

Separate natural-language guidance from a tool permission and from an operating-system boundary. Test one allowed action and one blocked action with synthetic data; retain the activity record and exact approval.

Minimal starting policy

1. Permit read-only work on a named test workspace.
2. Require explicit approval for a new recipient, purchase or irreversible deletion.
3. Set a stop condition and a retry ceiling for background tasks.

Evidence and limits

Official documentation checked

• Source-backed facts plus BotClaw editorial interpretation. No product account, inference or agent performance test was run.

Sources

• https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse
• https://docs.openclaw.ai/gateway/security
• https://hermes-agent.nousresearch.com/docs/user-guide/security
• https://help.openai.com/en/articles/20001530-getting-started-with-your-dot

Continue the decision

• Are two agents actually two isolated workspaces?
• Build a permission boundary that can be tested