™Every AI agent is built from a handful of parts. The first two are the model and the instructions.
The model: the thinking part
The model is the large language model underneath, the same kind of AI you met in the first course. It reads, reasons, decides what to do next and writes.
Different models have different strengths. Some are faster and cheaper, which suits simple, high-volume jobs like sorting emails. Others are slower and more capable, which suits jobs that need careful reasoning. Choosing a model is a trade-off between quality, speed and cost.
For most small business owners, the model choice matters less than people think. The instructions matter more.
The instructions: the job description
The instructions tell the agent who it is, what it's for, and how to behave. Technical people call this a system prompt. A more useful way to think about it is a job description for a new hire.
Good instructions usually cover:
- The role. "You handle first replies to new customer inquiries for a residential plumbing company."
- The goal. "Get every inquiry a helpful reply within minutes, and book a visit where it makes sense."
- What it knows. Your services, service area, hours and typical prices, or where to find them.
- The rules. "Never quote a final price. Never promise a same-day visit. Always offer two time slots."
- The tone. "Friendly, plain, short. Sign off as the office team."
- What to do when unsure. "If the request is an emergency or a complaint, pass it to a person immediately."
- The format. What a good output looks like, ideally with an example.
Why instructions make or break an agent
A capable model with vague instructions behaves like a talented new hire on day one with no training. It will do something reasonable, but probably not what you wanted.
The fix is the same as with people: be specific, give examples, write down the rules you'd normally explain out loud, and revise the instructions whenever you see it make a mistake.