How Claude actually gets things done

In Lesson 2.4 you watched a single turn go by: Claude thinks, reaches for a tool, reads the output, answers. That’s one step.

Most real work doesn’t fit in one step. So Claude does the obvious thing — it goes again.

An agent works in a loop: it makes a move, looks at what happened, and uses what it sees to choose the next move. Over and over, until the goal is met — or it stops to ask you.

That loop is the whole difference between a chatbot and an agent.

It’s a GPS, not a vending machine

The picture most people start with is a vending machine: drop in a request, a finished result falls out. One shot, done.

What Claude actually does is more like a GPS. It doesn’t compute your whole route once and then go quiet. It watches where you actually are, and the second you miss a turn, it reroutes — every instruction it gives is based on the last thing it saw.

Claude reroutes the same way. It changes a line, runs the thing, and looks. Still broken? It reads the new error, forms a fresh guess, tries again. Found the issue? Now it knows where to go next. Each move is chosen from the result of the last one — not from a script written in advance.

That’s why Claude can take on a task you couldn’t spell out perfectly up front. You don’t need to know every step. It doesn’t need to know every step. The loop finds them.

The most important beat is “observe”

Four beats keep coming back around: plan → act → observe → repeat. The quiet one — observe — is the one that matters most, and for two different people.

For Claude, observe is where it course-corrects. A move that didn’t work isn’t a dead end; it’s the information that shapes the next move.

For you, observe is your window in. Remember from Lesson 2.7 that Claude can be confidently wrong? The moment right after it acts — when the result is sitting there in the open — is exactly when a mistake is cheapest to catch. Watch the loop turn and you’re not trusting blindly; you’re checking, one move at a time.

Your turn — be the agent

Reading about a loop is one thing. Driving one is another. Below, you’re the agent: you get a small goal, and at each step you choose the next move and see what it reveals. Watch how the result of one move tells you what to do next.

What’s next

You’ve driven the loop from the inside. Time to run it for real: in your first useful task, you’ll point Claude at a real folder, ask for something, and watch a true plan → act → observe loop play out on your own machine.