How we think about agent safety

Live-data agents can do real damage if you let them. This is the set of constraints we enforce by default.

A live-data agent can do real damage. We treat that as a primary design constraint, not a footnote.

Default 1: scoped credentials. Every integration uses the smallest scope that lets the Buddy do its job. Read where possible; write only where the user explicitly authorises a workflow.

Default 2: human-in-the-loop for state-changing actions. Sending an email, posting to social, refunding an order — every one of these surfaces a confirmation step the first time, and remains explicit until the user opts in to autopilot.

Default 3: full audit trail. Every action a Buddy takes is logged with the prompt, tool call, parameters, and result. If something goes wrong you can read exactly what happened.