Reasoning / Research term
Reasoning and Acting (ReAct)
Reasoning and Acting (ReAct) is a 2023 agent pattern that alternates between choosing a next action, using a tool, reading the result, and choosing again.
Reasoning and Acting (ReAct) interleaves model decisions with observations from tools. For a flight search, an agent can request JFK results, read a tool error, and choose a bounded retry, another permitted search tool, or a clear failure report. The published pattern writes a reasoning step before each action, but current products often keep reasoning private and expose tool calls, tool results, and the final answer. ReAct remains a useful historical description of one control loop. Current agents may also use programmatic tool calling, a planner and executor, deterministic code, or multiple agents.
Builder example
ReAct's observable action-result sequence is useful for debugging, even when private reasoning is unavailable. Record each tool request, permission decision, response, state change, retry, and final evidence. A failure may come from the model, a misleading tool description, missing access, malformed data, application code, or a bad stopping rule. The log should identify which observable step failed.
Common confusion: Agent is the broad category. ReAct is one architecture inside that category. An application that calls several tools has not necessarily implemented the paper's explicit think-act-observe loop.