Approvals & safety

Every agent that sends to a customer, spends money, or changes production data queues a proposal in your Approvals inbox instead of acting silently. You review, edit, and approve — nothing sensitive ships without a human tap.

Read-only agents (analysis, monitoring, drafting) don't need approvals; they publish results and let you decide what to do next.

PostPilot's proposed posts. AdMiral's budget changes. ColdReach's outbound sends. ShopOps's price changes. RingPilot's outbound calls. WarrantyBot's refunds beyond a threshold. Every agent's manual lists exactly what it will and won't do unattended.

You can turn on rule-based auto-approval per agent (e.g. 'auto-post if the AI confidence is > 0.9 and it's a template we've approved before'). Auto-approval is off by default.

GuideFlow, PropertyCine, and any agent producing customer-facing safety instructions run through a citation and hallucination check before you see the output. If a claim can't be cited, the agent flags it instead of publishing.