The HITL contract
Every agent honours the same human-in-the-loop contract:- Read freely. Agents can read your data — campaign performance, deal history, social mentions, call transcripts, search rankings — without asking permission for each read.
- Propose, never execute. Any action that changes the world goes through the Inbox as a proposal. The agent does not write to Meta, send an email, dial a phone, or publish a post without your approval (or an explicit autonomy threshold you set).
- Surface attribution. Every proposal carries a “From: <agent>” badge so you always know which agent drafted what.
- Share through substrate. Agents talk to each other through three patterns: the Inbox (operator-mediated), the Brain (anonymised cross-brand wisdom), and Memory (brand-scoped facts). They never call each other directly.
What every agent will not do
- It will not become a nav destination. You navigate to the work surface for the domain, not to the agent.
- It will not write to another agent’s data store.
- It will not leak across brands. Every read and every action is scoped to a brand you’ve authorised the agent on.
- It will not surface in your Inbox without scoping itself to your authorised brand set.