The bottom rung is approve-everything: agents research, score, and draft freely, but every outward-facing action — an email, a reply, a post, money spent — lands in a queue for you to approve, edit, or dismiss. Every new agent team should start here, because the only honest way to judge an agent's judgment is to read what it actually proposes. After a week of reviewing drafts, you know precisely where an agent is reliable and where it still needs your eyes. That evidence, not optimism, is what justifies the next rung.
The middle rungs grant routine autonomy within limits: an agent handles a defined slice of work on its own, while guardrails like a daily send limit, a send window, and a do-not-contact list stay enforced and every action is still written to an audit log. The ladder runs in both directions — you can step an agent back down the moment something feels off, and a kill-switch stops everything instantly. Autonomy that can't be revoked isn't trust; it's a liability. The ladder turns delegation into a series of small, recoverable decisions instead of one large irreversible one.
In Brohns, the ladder is set per ecosystem rather than globally: your newest team can sit on approve-everything while a team you've watched for a few weeks runs independently within its limits, 24/7. The bottom rung isn't a UI promise but a server-enforced floor — the recipient and content of anything sent are read from the database at approval time, so a UI glitch or a clever prompt can't skip a rung you haven't granted. Climbing also gets easier over time, because agents learn from you: edit a draft before approving it, and the agent distills your correction into a lasting lesson. For a practical view of the first climb, see your first week with an AI agent team and the guide to approval-first AI.