Vercel added domain purchasing to its command-line tool. It built a pause into the flow for any agent using it. Before an agent completes a purchase, the tool returns a structured error. That error tells the agent to stop and ask a human to confirm. The human sees the domain name and the price, then decides yes or no. The agent proceeds or drops the request. No purchase happens without that explicit step. The design choice worth noting is where Vercel drew the line. Searching for domains, checking availability, and planning a deployment are all things an agent can do without interruption. Spending money requires a human decision. The tool enforces that boundary through an error format that agents can read and act on. It does not use a separate approval screen or a settings toggle. For a product team building agent features, this is a worked example of scope-splitting. You decide in advance which actions an agent can complete on its own, and which ones require the human to say yes. The split here follows money. The implementation is a machine-readable error that tells the agent what to do next. That keeps the agent in the loop without letting it past the spending step.