An Agent Needs a Boundary, Not Just a Task

A useful lesson in agent design: “handle the follow-up” is not a complete brief. It hides the real questions—what may the agent access, what may it do, and when must a person decide?

Imagine an agent that drafts follow-up messages from meeting notes. A safer first trial is to let it prepare drafts, show the source notes, and wait for human approval before sending. The test is not whether it produces fluent text; it’s whether the draft saves effort without introducing errors or crossing anyone’s consent boundary.

Start with one workflow, one clear permission limit, and a review point. If people can’t tell what the agent used or change its draft easily, pause before giving it more authority. That boundary is part of the product—not paperwork after the prototype.