ReAct Pattern

The value of interleaving is that reasoning is grounded in evidence the agent has actually gathered. A plan-then-execute agent commits to a full sequence before seeing any results, so a wrong assumption at step one corrupts everything after it. A ReAct-style loop re-plans after each observation, which makes it far more robust to a search returning nothing, an API shape differing from expectation, or a file not being where the model assumed.

The cost is latency and tokens. Every cycle is a full model call carrying the entire accumulated transcript, so a ten-step run is ten calls over a context that grows monotonically. Practical implementations therefore bound the loop with a step limit, compact or summarise older observations, and truncate verbose tool results before appending them. Without those controls the failure mode is not an error but a slow, expensive run that ends at the step ceiling with nothing to show.

The misconception is that ReAct is the right default because it is the best known. Many tasks have a known shape — retrieve, then summarise, then format — and encoding that shape directly as a pipeline is faster, cheaper and far easier to test than asking a model to rediscover it on every request. Reserve the open loop for genuinely variable work where the number and order of steps cannot be known in advance.