ReAct joined reasoning traces to environmental feedback
Shunyu Yao and coauthors introduced ReAct in a 2022 paper to combine two capabilities that had often been studied separately: reasoning over a task and acting to gather information. An action produces an observation, and that observation changes the next reasoning step.
ReAct is not a synonym for chain of thought or tool calling
Chain-of-thought prompting elicits intermediate reasoning but may never touch an external system. Tool calling supplies an action interface but does not define the surrounding decision loop.
Plan-and-Solve creates a plan before execution, whereas ReAct can revise its direction after each observation.
The trajectory matters more than a narrated thought
The inspectable ReAct structure is the sequence of action requests and returned observations. A free-form reasoning trace may sound explanatory, but it is not required to prove which source was read or which operation changed state.
Production systems can keep concise decision labels and structured state without exposing private reasoning text. What matters is whether each action follows from the available evidence and changes the task state.
An observation is useful only when it changes the next step
A ReAct loop should record the current question, the selected action, the returned observation, and the reason another action is needed. Repeating the same search with slightly different wording is not progress unless the new result resolves a named uncertainty.
A documentation lookup shows when the loop earns its cost
Suppose an agent must identify the current authentication parameter in an API. It searches the official reference, observes two versioned pages, checks the version in the user’s configuration, then answers from the matching page.
Each observation removes an ambiguity that the initial prompt could not resolve.
The agent-loop anatomy covers the production controls around this pattern. ReAct is unnecessary when one deterministic query returns the needed fact, and the common misuse is exposing verbose reasoning text as if it were an audit trail.
Tool inputs, observations, and state transitions are the inspectable evidence.
Stopping rules prevent a failed search from becoming a loop
The host can cap actions, detect repeated calls, reject unchanged state, and ask the user when required information is unavailable. Those controls belong outside the model because a prompt asking the model to stop responsibly is not an enforcement mechanism.
ReAct does not fit a task with a known fixed sequence and deterministic checks. In that case, a workflow makes the route visible and prevents the model from spending actions rediscovering it.
Continue with these glossary entries: