Explore

How an AI agent works, one piece at a time

Build an agent one piece at a time. Each piece fixes a problem and creates a new one. About 10 minutes.

Step 01 of 06

One model call

A customer types a question into a support chat. That customer is the user. The app is the software around the model: the chat window, plus the code behind it that sends text to the model and shows the reply. Later it also runs tools, checks permissions and keeps memory. The model only predicts text from what the app sends it, and never runs anything itself.

Your guess

The user asks about your company's refund policy. What happens?

Make a guess to continue.

The answer is B. The app sends the model only the question. The model has never seen your policy and has no way to look anything up, so it says it doesn't know or writes something that sounds right.

What can go wrong now

Asked about your own business, it guesses from what it learned in training, and the guess sounds right. Nothing tells the user it's a guess.

How teams handle it

Don't let it answer questions about your business without your own information in front of it, and have it say “I don't know” when that information is missing.

Next:To answer questions about your business, the model needs your own information in front of it: Add context.

Questions people ask

What is the app?

The software around the model: the chat window or support widget the user types into, plus the code behind it. It sends text to the model and shows the reply, and later it runs tools, checks permissions and keeps memory. Builders often call it the harness.

Is the agent the same as the model?

No. The model predicts text and never runs anything. The agent is the app and the model working together in a loop with tools: the model writes what to do next, and the app does it.