Introduction

ReAct (Reasoning + Acting) is a prompting technique that interleaves a model's step-by-step reasoning (as in Chain of Thought) with the ability to take concrete actions — like searching the web, querying a database, or calling a tool — and incorporate the results of those actions back into its ongoing reasoning process. Rather than reasoning purely from what it already knows internally, a ReAct-style model can pause mid-thought, retrieve fresh information from the outside world, and continue reasoning with that new information in hand.

ReAct directly addresses one of the core limitations discussed throughout earlier topics: a language model's knowledge is frozen at training time, and its reasoning — however careful — is still bounded by what it already learned. By combining reasoning with real actions, ReAct allows a model to ground its conclusions in current, verifiable information rather than relying solely on internal, potentially outdated or hallucination-prone knowledge.

Why Does ReAct Matter?

ReAct helps to:

• Ground a model's reasoning in real, current information rather than only internal knowledge
• Reduce hallucination risk by allowing verification through actual tool use mid-task
• Enable multi-step tasks that require both thinking and interacting with external systems
• Provide the conceptual foundation behind modern "agentic" AI applications
• Combine the interpretability of visible reasoning with the practical power of real actions
• Support tasks that simply cannot be completed through reasoning or knowledge alone

The Core ReAct Loop

Whiteboard

The ReAct pattern repeats a simple cycle:

Thought  → the model reasons about what it currently knows
           and what it still needs to find out
Action   → the model decides to use a specific tool to get
           that missing information
Observation → the result of that action is fed back into
           the model's context
(repeat) → the model reasons again, now with this new
           information available, deciding whether more
           actions are needed or it's ready to answer

A Simple Illustrative Example

User: "What's the current population of the capital city of
the country that won the most recent FIFA World Cup?"

Thought: I need to find out which country won the most recent
World Cup, since I don't have reliable information about
"most recent" given my training cutoff.
Action: search("most recent FIFA World Cup winner")
Observation: [search result indicating the winning country]

Thought: Now I need to find that country's capital city.
Action: search("capital of [winning country]")
Observation: [search result with the capital city name]

Thought: Now I need the current population of that capital.
Action: search("current population of [capital city]")
Observation: [search result with population figure]

Thought: I now have enough information to answer.
Final Answer: [Capital city]'s current population is
approximately [X], based on [country]'s recent World Cup win.

Notice how each action retrieves a specific missing piece of current information, and each observation directly informs the next reasoning step — something pure Chain of Thought, working from internal knowledge alone, couldn't reliably accomplish for a question involving current events.

Why ReAct Improves on Pure Reasoning Alone

Chain of Thought (covered in the previous topic) improves HOW
a model reasons through a problem, but it's still fundamentally
limited to whatever knowledge the model already has internally
— it can't look anything up.

ReAct extends this by giving the model a way to actually PAUSE
its reasoning, take a real action to retrieve new information,
and then CONTINUE reasoning with that fresh information now
available — directly addressing the knowledge-cutoff and
hallucination risks discussed in earlier LLM Behavior topics.

Common Tools Used in ReAct-Style Systems

ReAct and Modern Agentic AI

The ReAct pattern is the conceptual foundation behind what's
now commonly called "agentic" AI — systems that don't just
respond to a single prompt, but autonomously reason, take
actions, observe results, and continue working through
multi-step tasks with real tools, much like the searching and
tool-calling behavior a modern AI assistant performs when it
needs current information beyond its training data.

ReAct vs Chain of Thought

ReAct vs Self-Consistency

Key Properties of ReAct

• ReAct interleaves reasoning ("Thought"), tool use ("Action"), and results ("Observation") in a repeating cycle.
• It allows a model to retrieve current, external information mid-task, rather than relying solely on internal knowledge.
• The technique directly reduces hallucination risk for questions requiring up-to-date or verifiable information.
• ReAct is the conceptual foundation behind modern agentic AI systems that reason and take real actions.
• It can be combined with other techniques, like self-consistency, since they address different underlying problems.

Where Is ReAct Used?

Advantages

• Grounds model reasoning in current, verifiable, external information
• Directly reduces hallucination risk for time-sensitive or fact-dependent questions
• Enables genuinely multi-step, real-world task completion beyond what pure reasoning allows
• Makes the model's reasoning and information-gathering process transparent and inspectable
• Forms the practical foundation for building useful, real-world AI agents and assistants

Limitations

• Requires infrastructure to define, call, and handle results from external tools
• Adds latency, since each action requires waiting for a real result before continuing
• Errors can still occur in how the model interprets or acts on retrieved information
• More complex to implement and debug than pure text-based prompting techniques
• Tool availability and reliability directly limit what the model can actually accomplish

Real-World Examples

Best Practices

• Use ReAct specifically for tasks requiring current, external, or verifiable information beyond training data.
• Clearly define the available tools/actions and their expected inputs and outputs.
• Design prompts that explicitly encourage the Thought → Action → Observation cycle structure.
• Combine ReAct with self-consistency or careful verification for especially high-stakes tasks.
• Monitor and log the reasoning/action trace, since it provides valuable transparency into how a conclusion was reached.

Interview Tip

A common interview question is:

"What is the ReAct prompting technique, and how does it address limitations that Chain of Thought alone cannot?"

A strong answer is:

ReAct combines reasoning with the ability to take actions — like searching the web or querying a database — interleaving Thought, Action, and Observation steps in a repeating cycle, allowing a model to pause its reasoning, retrieve current external information, and then continue reasoning with that new information available. This addresses a fundamental limitation of Chain of Thought alone: pure reasoning, however careful, is still bounded by whatever knowledge the model already has internally from training, which becomes a problem for questions involving current events, real-time data, or anything requiring external verification — ReAct grounds the model's conclusions in real, retrievable information rather than relying solely on potentially outdated or hallucination-prone internal knowledge.

Explicitly naming the knowledge-cutoff limitation ReAct solves makes your answer stronger.

Conclusion

ReAct extends Chain of Thought by combining reasoning with real actions, allowing a model to retrieve current, external information mid-task and ground its conclusions in verifiable results rather than internal knowledge alone. This completes the full Core Techniques section — from Zero-Shot and Few-Shot prompting, through Role Prompting for tone and persona, to Chain of Thought and Self-Consistency for improved reasoning, and now ReAct for combining reasoning with real-world action — together forming the practical toolkit behind how modern AI assistants and agentic systems actually operate.