The Connected Agent

Where an agent stops writing words and starts changing records in systems other people rely on, logged as you, often with no undo. What MCP is, the three escalating capabilities and how to scope to the first, and what an OAuth consent screen actually grants. Then the half of the book that earns its place: how prompt injection really works, and why the hostile text is as likely to come from your own shared calendar as from a malicious server - built around a student whose agent was told, by a calendar entry, to email an attendee list to a stranger. Includes a seven-step checklist to run before connecting anything.