Modern AI agents, explained by the ReAct paper — three words and a loop

In October 2022, researchers from Princeton and Google Brain published a paper called ReAct: Synergizing Reasoning and Acting in Language Models. It is the ancestor of everything you now call an AI agent: ChatGPT tool use, Claude tool calling, coding agents. The name is literally Reasoning + Acting.

What is an agent?

A chatbot only talks. An agent makes a plan, actually uses tools, checks the results, and decides what to do next. It can search the web, run code, open files.

ReAct's answer is surprisingly simple: let the model think, act, and look at results — in a loop. No extra training needed; a few examples in the prompt were enough.

Three words to remember

Thought — talking to yourself

The agent's inner plan. "I need water, but first check the pot." The paper says this step grounds the answer and leads the next action.

Action — actually doing it

Calling a tool: Search("..."), Calculator("..."). This is where agents differ from plain chatbots.

Observation — the world answering back

Real information returning from the action: search results, code output. It feeds the next thought, and the loop goes around.

[Interactive 1] Ramen agent simulator

Watch an agent handle "Is there ramen? Cook it for me" step by step:

[Interactive 2] Chatbot vs ReAct agent

Same question, completely different process:

The whole loop

flowchart TD
  Q[User question] --> T1[Thought: what to check?]
  T1 --> A1[Action: call tool]
  A1 --> O1[Observation: results return]
  O1 --> T2{Enough?}
  T2 -->|no| T1
  T2 -->|yes| F[Final answer]

[Interactive 3] Sorting game

Is each sentence a Thought, an Action, or an Observation?

[Interactive 4] Final quiz

So what IS a modern agent?

After ReAct, tool calling became a built-in feature of major models. Strip any modern agent down and you get:

agent = LLM + tools + the ReAct loop
(think → act → observe, repeated until the goal is done)

The difference from 2022: back then prompts taught the loop; now models are trained to run it natively.

Sources

Paper: arXiv:2210.03629 — ReAct: Synergizing Reasoning and Acting in Language Models (Yao et al., 2022, presented at ICLR 2023)
Experiments: HotpotQA (multi-hop QA), FEVER (fact verification), ALFWorld, WebShop — improved performance and reduced hallucination.

★ PINION ★