🧭 The skeleton: reason → act → observe

One agent loop is just an AI that reasons what to do, acts, e observes the result—again and again, until the goal is reached. Nothing magical: three beats that repeat toward a goal, until a "done?" check says to stop.

Watch five videos and you’ll see five different figures. They’re the same concept from different angles. These are the four you’ll encounter most often—and the confusion isn’t your fault.

What to look at: four drawings, just one plant. Beginners think they’re four different things; they’re four pictures of the same object.

Of the three beats, the middle one in the name — observe — this is what makes the loop actually work. It’s the moment when the agent reads its own result instead of assuming it worked.

What to look at: the agent reads the result, and the information feeds into the next decision. Without that read, it’s blind.

✗ No observation

The agent acts, assumes it worked, and moves on. The result: polished, confident, and wrong — and it calls it “done.”

✓ Observing

The agent runs the test, reads the actual output, notices what's still missing—and adjusts. That's what makes the loop converge instead of deriving.

The best mental model: think of a a smart intern you don’t micromanage. You provide the goal. It figures out the next step, does it, checks its own work, and continues — coming back to you only when it’s done (or gets stuck).

What to look at: it works alone and checks its own checklist. You don’t dictate every step—you give it the goal and the definition of "done."

1

You deliver the goal, not the list of steps.

2

It figures out the next step and runs.

3

It checks its own work and repeats.

4

It comes back to you only when it finishes or gets stuck.

AI never gets it right on the first try. In a single prompt (one-shot) you accept the answer from the first attempt. In a loop, AI takes on the role of the human who gives feedback and iterates—and that's why it gets much closer.

To wrap up the module: each turn is a step, and the agent cycles—act, observe, act, observe—until the “done?” check passes, or a guardrail (a safety limit, such as a maximum number of attempts) stop. You can already get a feel for a real loop running.

# Cole no Claude Code — seu primeiríssimo loop (você vai aprofundar na Trilha 3)

# o objetivo
Conserte o único teste que está falhando neste projeto.

# o que "pronto" significa (deixe checável)
Pronto quando: esse teste passa ao rodar a suíte de testes.

# como checar
Verifique rodando os testes e lendo a saída.

# o guardrail
Continue até passar, ou pare depois de 5 tentativas e me avise.

How to verify: open a small project with 1 failing test, paste the prompt into Claude Code, and watch it cycle through reason→act→observe until the tests pass (or it stops after 5 attempts). That's the skeleton of this module, running.