The word has two lives. In reinforcement learning, an agent is any system that observes an environment, chooses actions according to a policy and receives rewards: the Atari player, or AlphaGo. In recent jargon, an agent is a system that uses a language model to decide which actions to take and in what order, with two key components: tools to do what the model cannot do alone (tool use) and a strategy for planning (Alammar & Grootendorst, 2024, p. 218).
The most influential scheme is ReAct, which interleaves reasoning and acting (Yao et al., 2023). The model repeats a three-step cycle (Alammar & Grootendorst, 2024, p. 219):
Thought: I need the current price in dollars and then convert it to euros. Action: search["MacBook Pro price"] → Observation: a list of prices in USD… Thought: now I use the calculator with the exchange rate…
On questions that require looking things up in Wikipedia, ReAct reduced the hallucination and error propagation of pure chain of thought reasoning (Yao et al., 2023). Since 2025, the most visible variant has been coding agents, which read a repository, edit files and run tests over many steps (LLM agent).
Are they agents in the philosophical sense? Millière & Buckner (2024b) raise the question: they are described as systems that generate, store and break down goals, yet they have no intrinsic goals or motivations; whether the word is a misnomer depends on how thick one's notion of agency is (2024b, sec. 3.1). From the side of intentionality, Dennett (1987) offers a tempting answer: if adopting the intentional stance toward the system predicts its behavior well, that is all there is to having beliefs and desires. Skeptics reply that the goal really lives in the prompt somebody else wrote.