Most people pick an AI agent by what the task is called. That's the wrong axis. Pick by what the agent can physically reach.
A terminal agent owns disk, git, packages and CLIs. A managed app-builder owns the frontend, the backend and the deploy. A browser agent owns any scriptable web UI — dashboards, settings pages, forms. Direct database and API tools own reads against live systems. And you own only the irreducible: real-browser sign-ins, the ship decision, the visual judgment call. One agent stays the architect — it plans, writes the briefs the others execute, and diagnoses. It does not try to be the hands.
Get this wrong and you get wrong-layer work: an agent that can't see a rendered page is asked whether the page looks right, so it guesses. Run the seed: take a task and assign each step to the layer that can actually reach it.