From "answer a question" to "achieve a goal"
An ordinary LLM call is "question to answer". An agent is the next level: an LLM given a goal, which decides itself what to do, uses tools, checks results and repeats steps until it reaches the goal. The difference is like "give me a fact" vs "solve the task by taking the needed steps".
The key distinction of an agent
- A goal, not a one-off answer: "handle this", not "answer this".
- Actions through tools: the agent calls functions, gets data, performs steps.
- Iteration: think to act to see the result to adjust to repeat.
- Autonomy (partial): it decides the sequence of steps to the goal itself.
What it looks like
Explain the difference between an ordinary LLM call and an agent with an example.
Example: the task "find out which of our products sold worst last month and prepare a draft email to the supplier". An ordinary LLM: you ask a question — it answers with text based on what it knows or what you gave in the prompt. But it does not know the sales data, will not send the email, does not "do" anything itself. An agent works differently: 1) Gets a GOAL, not a question. 2) Thinks: "To find the worst product I need sales data — I will call the get_sales_data tool". 3) Acts: calls the tool, gets the data. 4) Sees the result: analyzes, finds the worst product. 5) Thinks further: "Now I need to prepare an email — I will compose a draft". 6) Acts: generates the draft. 7) Checks: is the goal reached? Yes — returns the result (or, for the important, hands it to a human for confirmation). Key: the agent builds the SEQUENCE of steps to the goal itself, using tools and adjusting along the way, rather than just giving one answer. It is a cycle of "thought to did to checked to repeated" until the goal is reached. It is precisely this iterative work with tools that turns an LLM from a "conversationalist" into a "task executor". But with this power comes responsibility (safety, control) — which has its own module.
The agent cycle: think — act — observe
At the core of an agent is a cycle: the agent thinks (what to do next for the goal), acts (calls a tool), observes the result, and based on it thinks again — until the goal is reached. This cycle (often called "reasoning + acting") lets the agent solve multi-step tasks rather than only answering in one go. Understanding this cycle is the basis of the whole course.
Course rule: an agent is an LLM that, to reach a goal, iteratively thinks, acts with tools and observes the result. This is the shift from "an answer" to "solving a task in steps".
🧠 How does an agent differ from an ordinary LLM call?