ReAct

A model asked to reason produces a plausible chain and no way to check it. A model asked only to act has no plan. ReAct alternates: a thought, then an action, then the result of that action, which the next thought gets to read. The world's answer, rather than the model's confidence, is what corrects it.

Thought, action, observation, repeated until the model emits an answer or a step limit stops it. The value is that an observation is evidence the model did not generate, so a wrong assumption gets contradicted by a real result instead of compounding quietly. The failure mode is the loop that never converges, which is why the step limit is not optional.

Step n is conditioned on every thought, action and observation before it, so the prompt is Θ(n) tokens by step n and attention over it is Θ(n²). A long run is quadratically expensive in its own history — which is why trimming or summarising the trajectory becomes necessary rather than merely tidy.