Permission lanes for AI agents

Agents are fast. That is the problem. A permission lane is a short written rule for one kind of action.

A lane says:

• Who may act (which agent or role).
• What it may do on its own, and what always goes to a human.
• How often (a daily cap, one follow-up at most).
• Where every action is logged.

Good defaults: money, refunds, sign-ins, legal wording and anything sent to a new audience stay with a human. Replies to people who wrote in first, using approved facts, can often run on their own.

Start with one lane, watch the log for a week, then widen it. A lane you never review is just a hope.