ReAct (Reasoning and Acting)
An iterative cognitive architecture alternating reasoning, action and observation, the basis of many AI agents.
ICLR 2023: ReAct, by Shunyu Yao and coauthors from Princeton and Google Research, beat imitation and reinforcement learning methods by 34 points of success rate on ALFWorld and 10 points on WebShop. It needed only one or two examples in the prompt. ReAct, short for Reasoning and Acting, is an iterative cognitive architecture that Yao and coauthors formalized in 2022, and one of the technical foundations of agentic AI. Instead of answering in a single pass, the model alternates between steps. In reasoning steps it analyzes the problem and plans; in action steps it calls an external tool such as a search or a query; in observation steps it reads the result before moving on. The loop repeats until the goal is reached, and the model corrects itself when an observation reveals an error. ReAct combines internal planning with interaction with the outside world, so the model bases its decisions on up-to-date information rather than on its training memory alone, which cuts down certain hallucinations. For understanding risk, ReAct shows concretely how a large language model becomes an agent able to act autonomously and therefore to set real consequences in motion. This chaining autonomy, with no human checkpoint between steps, is what creates the tight coupling of agentic systems and exposes them to rapid error propagation.
Textbook case: a cyber claims manager deploys a ReAct agent, a method published in October 2022, to prepare claim files. The agent receives an insured's notification and reasons that it must first check policy eligibility. It queries the policy-management database and sees that the premium is paid. It then reasons that 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, which a handler validates in ten minutes instead of spending two hours on it.
ReAct, reasoning and acting, ReAct loop
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.