The ReAct pattern: think, act, observe, repeat

How the loop works

On its own, a large language model only produces text. ReAct gives it a three-beat routine that repeats until the job is done:

1. Thought: the model writes out what it has understood and what it intends to do.
2. Action: it picks a tool (web search, calculator, file reader) and its arguments, through tool calling.
3. Observation: the surrounding program runs the action and feeds the result back, where it joins the model’s context.

The loop ends when the model decides it has enough to answer, or when a limit set by the developer kicks in.

A worked example

Question: “What is the current population of the town where Ernest Hemingway was born?”

• Thought: “First I need Hemingway’s birthplace.” Action: search “Ernest Hemingway birthplace”. Observation: Oak Park, Illinois.
• Thought: “Now I need the latest population figure for Oak Park.” Action: another search. Observation: an official census page.
• Thought: “The page gives a figure and a date; I can answer and cite it.” End of loop.

Without ReAct, the model would answer in one go from memory, risking a hallucination or out-of-date information. With ReAct, each step rests on an observation you can check.

Where it came from

The pattern was introduced in the paper “ReAct: Synergizing Reasoning and Acting in Language Models” (Shunyu Yao et al., October 2022, published at ICLR 2023). It merged two lines of work that had been separate: chain of thought, where a model reasons without touching the outside world, and models that act without spelling out their reasoning. The authors showed that interleaving the two makes answers more reliable and traces easier to read.

Why it matters, and where it falls short

ReAct has become the implicit skeleton of many AI agents and frameworks: even when the name never appears, the “think, call a tool, read the result” loop is there. Its weaknesses are well known. Every turn re-reads the whole history, so token costs climb quickly. One bad observation can send the agent down the wrong path. And an agent that does not know when to stop can go round in circles. That is why variants plan first (task decomposition) or add a self-reflection step.

Common myths

• “The visible reasoning truly explains the decision.” It helps with debugging, but nothing guarantees it faithfully mirrors the model’s internal computation.
• “More iterations are always better.” Past a certain point the agent piles noise into its context and costs more without getting more accurate.