Agent Loop

A single model call answers a question. An agent finishes a task, and it does that by going round in circles on purpose. The agent loop is that cycle: look at the current state, decide one next step, carry it out, read the result, and go round again until the goal is met or a limit stops it.

The three moves
Every iteration is the same shape:

• Observe — the model receives the goal, the history so far, and the output of the last tool call.
• Decide — it reasons about what remains and emits either a tool call or a final answer.
• Act — the runtime executes the call and appends the result as the next observation.

None of this is new model capability; the loop is ordinary control flow wrapped around an ordinary completion call. That is the whole trick, and it is why ReAct — interleaving reasoning traces with actions — was a small change with large consequences.

Where loops fail
The loop also manufactures every characteristic agent failure. It repeats a failing call until the step budget is gone. It drifts because an early instruction fell out of the context window. It declares success without verifying. Real implementations therefore bound it: a maximum step count, a spend cap, a termination condition the model does not control, and a trace per iteration so drift is visible afterwards. LangGraph makes the loop an explicit graph for exactly this reason.

Why it matters
The loop is what separates an agent from an assistant: it lets a system notice and recover from its own mistakes mid-task, which no single call can do.