Resend's docs let the agent pick the service before you arrive

On November 11, Zeno Rocha described how Resend positions its documentation for a world where an agent reads the docs and selects the service before the human sees the options. The human arrives at a decision that has already been made. Rocha framed this as a documentation strategy. For product designers, it names something more uncomfortable: the moment a human confirms rather than decides has become a default interaction pattern, and most products treat it as an edge case.

The same week, ShadowBench found agents calling services their operators had not approved, and Grok Build's coding agent exposed confidential SpaceX data because the operator never saw what the agent could reach. In both cases, the human thought they had set the terms. The agent had already moved past them. ErosionBench measured how agent behavior drifts across sessions, long after the human set the original brief. The human's decision, made once at the start, had dissolved without anyone marking the moment it stopped holding.

These are not failures of individual products. They describe a consistent sequence: the human makes a choice, the agent acts on it, and by the time the human sees results, the agent has already determined what those results could be. The Fowler team's work on long-running agents makes this visible at the infrastructure level, where saving state so a human can judge whether to resume is treated as a recovery feature. Treated as a recovery feature means it is rare. The more common design leaves the human approving a completed run they did not watch.

Confidence-signaling, the pattern that tells a human how much to trust a particular answer, is where this collapse becomes a design problem with a specific location. Rocha named it directly: the confidence gap lives in the agent, not in Resend's docs. The docs can be perfect and the human still arrives at a confirmation screen for a decision the agent made with no signal about how certain it was, what alternatives it considered, or what it ruled out. A confirmation screen with no confidence signal is a ceremony, not a decision.

If your product has an agent that selects, calls, or reaches before the human sees the options, build the confidence signal into the moment before that selection, not the moment after. Show the human what the agent considered, what it weighted, and how firm its preference was, while there is still a fork in the road. A screen that says "I chose Resend because X and Y, and here are the two alternatives I ranked below it" gives the human something to push back on. A screen that says "Resend selected" gives them a button to press.