Chat is one-shot Q&A. Agents (and IDE agent modes) try multi-step loops: read files, edit, run commands, adjust. They still hallucinate and can edit the wrong place.
You own: clear goal (DoD), clear bounds (don’t touch X; no secrets), clear acceptance (how we know it’s done).
Tie-in: practice explain→minimal fix in Learn Coding before letting an agent roam the repo.
Practice: contrast one-shot chat patches vs multi-step agent edits in three sentences, then run the Learn Coding scene.