ReAct

ReAct (Reasoning + Acting) is a prompting paradigm introduced by Yao et al. (2022) that fundamentally shaped how we build AI agents. It demonstrated that interleaving reasoning traces with actions produces more capable, interpretable, and reliable agent behavior.

The Core Insight

Before ReAct, there were two separate approaches:

• Chain-of-Thought (CoT): Models reason step-by-step, improving accuracy on complex tasks—but they can’t do anything
• Action-only agents: Models take actions in an environment—but lack transparent reasoning about why

ReAct unified these approaches. The key insight: reasoning and acting are synergistic.

The ReAct Loop

The pattern follows a simple structure:

graph TD
  T1["Thought<br/>need population data for Tokyo"] --> A1["Action<br/>search('Tokyo population 2024')"]
  A1 --> O1["Observation<br/>Tokyo: 13.96 million"]
  O1 --> T2["Thought<br/>have answer, can respond now"]
  T2 --> D{continue?}
  D -.yes.-> T1
  D -.no.-> F[finish]

  style T1 fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc
  style A1 fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc
  style O1 fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc
  style T2 fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc
  style D fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc
  style F fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc

Each iteration produces:

1. Thought: Explicit reasoning about the current state and next steps
2. Action: A concrete step taken in the environment
3. Observation: The result of that action, fed back into the loop

Example Trace

A typical ReAct trace looks like this:

Question: What is the elevation of the capital of the country
where Einstein was born?

Thought 1: I need to find where Einstein was born.
Action 1: search("Albert Einstein birthplace")
Observation 1: Albert Einstein was born in Ulm, Germany.

Thought 2: Einstein was born in Germany. The capital of Germany
is Berlin. I need to find Berlin's elevation.
Action 2: search("Berlin elevation")
Observation 2: Berlin has an average elevation of 34 meters.

Thought 3: I now have enough information to answer the question.
Action 3: finish("34 meters")

Why ReAct Matters

1. Interpretability

Every decision is justified in the thought trace. You can see why the agent did what it did.

2. Grounding

Actions provide real-world feedback that keeps reasoning on track. The model can’t drift into pure confabulation when observations contradict its assumptions.

3. Composability

The pattern works with any set of tools. Swap out the action space, and the same architecture applies.

4. Error Recovery

When an action fails or returns unexpected results, the thought step allows the agent to reason about the failure and try a different approach.

Variations and Extensions

Since the original paper, ReAct has spawned many variations:

Implementation Notes

When implementing ReAct, key considerations include:

• Prompt design: The format of thought/action/observation matters. Consistent formatting helps the model follow the pattern.
• Action space: Define clear, unambiguous actions. Vague actions lead to vague behavior.
• Observation parsing: Truncate or summarize long observations to stay within context limits.
• Loop limits: Set maximum iterations to prevent infinite loops.

The Legacy

ReAct didn’t invent agent loops—those existed before. What it did was formalize them, providing a template that could be studied, replicated, and improved upon.

Today, most agent frameworks implicitly or explicitly implement some variant of ReAct. It’s become the default grammar for agent behavior.

See Also

• The Agent Loop — the broader pattern ReAct instantiates
• Tool Use — the mechanics of agent actions
• Agentogenesis — the historical context for ReAct’s emergence