What Is an Agent?

Think about how you solve a task that you've never seen. You have a look, you try something, you see what happens, and then you decide what to do next. You don't plan every step up front. You loop. An agent does exactly the same.

• You look
• You try something
• You see what happens, then loop

The agent loop

Here it is. Think, act, observe. The model thinks about what to do. It acts, by calling a tool. It observes the result. And that result goes back into the next thought. This repeats until the goal is met, or until it gives up. Every agent product that you've ever seen, however impressive, runs on this loop.

The loop in twelve lines

Let's write it. This has no AI in it at all, just the shape. A list of notes holds everything so far. In a loop, we think, and we check whether we're done. If not, we act, and we add the result to our notes. And there's a maximum number of steps, so it can't run forever. Run it. There's the loop, working through its steps. Keep this shape in mind. Everything else in this course plugs into it.

def agent(goal, think, act, max_steps=5):
    notes = [goal]
    for step in range(1, max_steps + 1):
        decision = think(notes)
        print(f"Step {step}: {decision}")
        if decision == "done":
            return notes
        notes.append(act(decision))
    return notes

plan = iter(["look up order", "check policy", "done"])
agent("Help with order A17", lambda notes: next(plan), lambda d: f"result of {d}")

Step 1: look up order
Step 2: check policy
Step 3: done

The four parts of any agent

Every agent has four parts. A model, which does the thinking. Tools, which let it act. Memory, which holds what it knows. And instructions, which tell it its goal and rules. Change any one of those, and you've got a different agent. That's also the outline of this course.

Summary