The Agent Loop

Learn one model and the rest of agentic engineering stops being mysterious: an agent is a model running tools in a loop toward a goal. It takes a step, looks at the result, decides the next step, and repeats until the work is done.

Three moves, on repeat

Every turn of the loop is the same three moves:

• Think — the model reasons about what to do next.
• Act — it calls a tool: read a file, run a command, search the web.
• Observe — the result comes back, and the model reads it before deciding again.

That is the whole engine. A chatbot does one think-and-answer. An agent wraps that in a loop, so it can work through a real task instead of describing one.

Why it keeps going — and when it stops

The loop runs until a stop condition is met: the goal is reached, or a limit kicks in — a turn cap, a time budget, a human saying "stop." Take the stop condition away and a confused agent will loop until it runs out of room. The limit is a feature, not a flaw.

Step the loop yourself, one turn at a time, and watch the working memory fill as it goes:

What the loop explains

Once you see the loop, a lot of agent behavior makes sense:

• It can only see what is in front of it. If a step's result is not fed back in, the agent is blind to it — and will repeat or contradict itself.
• Longer tasks cost more, fast. Every turn re-reads the whole history, so a 40-step task pays for the early steps many times over.
• The control points are between the steps. You cannot edit the model's thinking, but you can decide which tools it is allowed to run and review what it did.

Where the picture gets fuzzy

"Think" is a useful word, not a literal one. The model is not deliberating the way you do; it is predicting a likely next step from patterns. The loop is real and mechanical — the "reasoning" inside each step is generated, not pondered. Hold both ideas at once and you will read agent behavior correctly.

Ready for the machinery? The Harness pillar builds this loop from the floor up, including the agent loop in full.