The idea is delightfully simple: instead of making the model only reason (CoT) or only act (like a plain tool-using agent), you interleave both.

The model produces reasoning traces ("here's what I'm thinking and why") and actions ("here's what I'm going to do about it") in an alternating loop.

The reasoning helps it plan, track progress, and recover from mistakes.

The acting lets it reach out to the real world i.e search engines, knowledge bases, environments and pull in fresh facts.

In short: reasoning decides what to look up next, and the retrieved info grounds the next round of reasoning.

They feed each other. It's less "lonely genius monologue" and more "detective who actually checks the evidence."

The result? On a bunch of language and decision-making tasks, ReAct beats several strong baselines, and as a bonus it's way more interpretable, you can literally read the model's train of thought and see why it did what it did.

The authors found the sweet spot is ReAct combined with CoT, so the model can lean on its own internal knowledge and external info when it needs to.

How It Actually Works

ReAct is inspired by something very human: we learn and make decisions by bouncing between thinking and doing.

You don't plan an entire road trip in your head and then drive it blindfolded you think, act, observe what happens, adjust, repeat.

The loop has three moving parts:

• 
  Thought: the model reasons about the current state and what to do next.


• 
  Action: the model does something (e.g., Search[...], Lookup[...], Finish[...]).


• 
  Observation: the environment responds with new info, which feeds the next Thought.

Here's the classic example from the paper, answering a question from HotpotQA:

Aside from the Apple Remote, what other devices can control the program Apple Remote was originally designed to interact with?

The headline: ReAct generally beats Act (acting only, no thinking) on both tasks, turns out a little reasoning goes a long way.

Against CoT it's more of a split decision:

• ReAct wins on Fever.


• ReAct lags slightly behind CoT on HotpotQA.

The paper digs into why, and the short version is a neat little trade-off:

• 
  CoT hallucinates facts: confident, fluent, occasionally fictional.


• 
  ReAct's rigid Thought-Act-Obs structure can box in its reasoning flexibility.


• 
  ReAct also leans hard on what it retrieves, feed it a junk search result and it can get derailed and struggle to recover.

The best of both worlds? Methods that let the model switch between ReAct and CoT + Self-Consistency outperform everything else.

Grounded when it needs facts, free-flowing when it needs to reason.

Have your cake, retrieve it too.

Results on Decision-Making Tasks

ReAct isn't just a trivia champ, it also shows up for interactive, action-driven tasks.

The paper evaluates it on two benchmarks: ALFWorld (a text-based game) and WebShop (a simulated online shopping environment).

Both throw the model into messy environments where it has to reason in order to act and explore effectively.

. All the images are from the above paper.

Disclaimer: This article was written by me; AI was used to fix grammar and improve readability.

Cover Image Credits:

AI agents write code fast. They also silently remove logic, change behavior, and introduce bugs — without telling you. You often find out in production.

git-lrc fixes this. It hooks into git commit and reviews every diff before it lands. 60-second setup. Completely free.

Any feedback or contributors are welcome! It's online, source-available, and ready for anyone to use.

⭐ Star it on GitHub: