Hold the two halves together. Agents are genuinely powerful: they act in the world, not just talk about it, and they shine on tasks with clear steps, checkable work, and recoverable mistakes. And agents are genuinely fallible: errors compound across a loop, confidence never signals correctness, and what they do can drift from what you meant, with consequences that leave the screen.

The honest response to both being true at once is not fear and not blind trust. It is the same calibrated trust from last module, now with real stakes: give an agent as much leash as the task can safely afford, and watch most closely where a mistake would cost the most.

That brings us to the end of the course, and to the questions that do not have clean answers yet. The last chapter sits with them honestly.