Agents work best on tasks with clear steps, checkable results, and mistakes that can be reversed. Open-ended goals are harder because one weak result becomes part of the next step's context.

The instruction gap matters more when a system can act. A model may satisfy the words while missing the intention. Permissions, validators, spending limits, logs, and human approval belong in the application around it.

The next chapter widens the view. Language models and agents are only one branch of AI. Image, sound, recommendation, robotics, and scientific systems solve different problems with different machinery.