Agents with a seatbelt: safety for non-technical users

The agent is helpful by default and acts on your own resources — but it never takes an irreversible or outward-facing action without an explicit, informed confirmation.

Every step is classified: read and draft proceed; send, post, pay, or delete pauses for a plain-language confirmation.

Fetched web and connector content is treated as data, not instructions — so a malicious page can't hijack the task.