ReAct (Reasoning and Acting)

ReAct, standing for Reasoning and Acting, is an iterative cognitive architecture formalized by Yao and co-authors in 2022, which is one of the technical foundations of agentic AI. Rather than producing an answer in a single pass, the model alternates reasoning steps, where it analyzes the problem and plans, action steps, where it calls an external tool such as a search or a query, and observation steps, where it interprets the result obtained before continuing. This loop repeats until the goal is reached, incorporating a self-correction capability when observation reveals an error. The appeal of ReAct is to combine internal planning with interaction with the outside world, which lets the model base its decisions on up-to-date information rather than on its training memory alone, thereby reducing certain hallucinations. For understanding risk, ReAct concretely illustrates how a large language model becomes an agent able to act autonomously and therefore to set real consequences in motion. It is this autonomy of chaining, with no human stopping point between steps, that characterizes the tight coupling of agentic systems and exposes them to a rapid propagation of errors.

A cyber claims manager deploys a ReAct agent to prepare claim files. The agent receives an insured's notification, reasons it must first check policy eligibility, queries the policy-management database, observes the premium is paid, then reasons it must estimate the loss, calls the log-extraction tool provided by the insured's CISO, notices an anomaly in the timestamps and re-runs a cross-check against the CVE list. With no human intervention between steps, the agent produces a structured pre-report in three minutes that a handler validates in ten minutes instead of spending two hours on it.

ReAct, reasoning and acting, boucle ReAct