LLMs & agents

The AI tutor in your panel is a different species from the racing CNN: a large language model wrapped into an agent.

What an LLM is

A neural network trained on enormous amounts of text to predict the next token. That single skill, at scale, yields something remarkable: you steer it with language. Your instructions are not code — they are text the model continues coherently.

• No task-specific training: describe the task, get the behavior.
• The flip side: it predicts plausible text. It can be confidently wrong, so agents must check the world instead of assuming.

From LLM to agent

An LLM alone can only talk. An agent = LLM + tools + a loop:

1. The model receives your request and a list of tools (name, description, parameters).
2. It replies either with text — or with a tool call.
3. The platform executes the call (read camera, drive, edit file) and feeds the result back.
4. Repeat until the task is done, then report.

You have watched this loop live in the agent course — every call is printed in the chat.

Why instructions and tools are the levers

• Instructions set identity, rules, and knowledge — they shape every prediction. Vague rules lose to concrete requests; specific rules win (practice).
• Tools define the agent's action space — it cannot do what it has no tool for, and it will try whatever tools you give it (practice).

Deeper mechanics of a single call → LLM agents in detail.