Zenaique

What did the Yao et al. 2022 ReAct paper establish about combining reasoning and acting?

Flashcard·Medium·4.0 · 0·~30s·Asked atAutodeskCopy AiTencent·Relevant atAnthropicLangChain
Attempt it
TL;DR

ReAct showed that interleaving reasoning traces with tool actions beats reasoning alone or acting alone, and it gave agents their Thought, Action, Observation loop.

Memory aid
Sign in to see the mnemonic that makes this stick.
Easy to grasp

Imagine doing a treasure hunt. One friend just thinks hard in their head about where the treasure is but never checks a single clue, so their confident guesses drift away from reality. Another friend runs around grabbing every clue but never stops to reason about what the clues mean, so they wander aimlessly. The smart approach is to do both, turn by turn: think a little about your next move, go check one clue, look at what you found, then think again with that new fact in hand. ReAct is the research that showed a language model works the same way. When it alternates between reasoning out loud and taking real actions like a search, it stays grounded in facts and plans better than doing either one on its own.

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

ReAct is one of the most cited papers in the modern agent literature, and for good reason. Published in 2022, its name compresses reasoning and acting, and its central claim is deceptively simple. A language model that interleaves reasoning steps with tool-use actions outperforms a model that only reasons or only acts.

That claim matters because it resolved a tension that had been building. Chain-of-thought had shown that letting a model think step by step improves multi-step reasoning. Separately, tool-use work had shown that a model can call a search engine or a calculator. ReAct asked what happens when you fuse the two into a single tight loop, and it answered with a clean experiment and a structure that the whole field then adopted.

What follows traces the experiment that made the case, the feedback mechanism that explains why the combination wins, the loop structure that became the default skeleton for agents, and the limitations that motivated the work built on top of it. The thread tying all of this together is that a small structural idea, repeated reliably, scaled into the foundation of an entire field.

The three-way comparison that made the point

ReAct's persuasive power comes from how cleanly it isolated the variable. The paper pitted three conditions against each other on the same benchmarks. The first was reasoning only, essentially chain-of-thought, where the model thinks step by step but never queries the outside world. The second was acting only, where the model issues tool calls but produces no explicit reasoning between them. The third was ReAct itself, alternating a reasoning step with an action and then reading the result.

On knowledge intensive tasks such as HotpotQA and Fever, reasoning only models hallucinated. They produced fluent, confident chains that drifted away from the facts because nothing ever checked a claim against a real source. On interactive decision tasks such as ALFWorld and WebShop, acting only agents floundered. They could push buttons and navigate, but with no plan they got lost and repeated mistakes.

ReAct beat both. The reasoning steps gave the agent a plan and let it decide which action made sense next. The actions, and the observations they returned, kept the reasoning anchored to reality. The two halves covered each other's blind spots, and the combined system was both more accurate and more interpretable.

Why interleaving works: grounding and feedback
The loop it bequeathed to every agent
Limits and what came after
Why interviewers treat ReAct as a landmark
Sign in to unlock the full deep dive.

Situations where this technique stops working.

Sign in to see when this approach fails.

2–4 min · Everything important, quickly.

Sign in to see the quick scan of the deep dive.

Real products, models, and research that use this idea.

  • LangGraph ships a prebuilt ReAct agent that wraps Anthropic or OpenAI function calling in a state graph, with each node running one Thought, Action, Observation turn.
  • Anthropic's tool use in Claude Opus 4.7 is ReAct under the hood: the model reasons, emits a typed tool call, reads the result, and continues until it answers.
Sign in to see more production examples.

What an interviewer would ask next. Try answering before peeking at the approach.

QHow does modern function calling change ReAct's original free text action format, and what problem does that solve?
A

Discuss how typed JSON schemas replace parsing actions out of prose, cutting malformed tool calls. Note the reasoning trace survives as scratchpad or thinking tokens while the action becomes a validated structured call.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Calling ReAct merely a prompting trick. Its lasting contribution is the Thought, Action, Observation loop that grounds reasoning in real tool results and underpins modern agents.

Sign in to see all red flags and common mistakes.

60 second bullets to scan on the way to the call.

  • State what ReAct stands for and the year it appeared.

  • Name the three conditions the paper compared and which won.

Sign in to unlock the revision sheet.

Primary sources. Browse if you want the original framing.

Similar questions

Same topic, related formats. Practice these next.

4 curated
Next question
What is the Model Context Protocol (MCP) and what problem does it solve?
MCQ·Easy