What is a ReAct agent?
A ReAct agent is a language model agent that interleaves written reasoning steps with actions such as tool calls, reading each observation before deciding the next step, following the ReAct method published by Yao and colleagues in 2022.
ReAct, short for reasoning and acting, comes from the paper "ReAct: Synergizing Reasoning and Acting in Language Models" by Shunyu Yao and co-authors, first posted in October 2022 and published at ICLR 2023. The idea was to have a model produce reasoning traces and task actions in an interleaved way: the reasoning helps it plan and revise, and the actions let it pull in information from outside.
In the paper, ReAct beat imitation and reinforcement learning baselines by an absolute 34% success rate on ALFWorld and 10% on WebShop, and reduced the hallucination seen in chain-of-thought prompting on question answering by looking facts up through a simple Wikipedia API.
How a ReAct agent works
Thought
The model writes a short reasoning step: what it knows, what it still needs and what to try next.
Action
It names a tool and the input to give it, such as a search query or a lookup key.
Observation
The harness runs the action and adds the result to the transcript.
Repeat or answer
The model reads the observation and either takes another step or writes the final answer.
ReAct in 2026
The original method parsed actions out of free text. Current models do the same thing through structured tool calling and built-in reasoning, so most agents now run a ReAct-style loop without calling it that. Google's whitepaper on agents describes the orchestration layer as a reason, act and observe cycle, and Google's Cloud Architecture Center still lists ReAct as a named design pattern.
The name has also faded from framework APIs. LangGraph 1.0 deprecated its create_react_agent helper, and LangChain's create_agent, which runs on LangGraph, took its place.
Where ReAct falls short
Cost per step
Every thought and action is a model call, so a long task becomes a long and expensive transcript.
Compounding errors
A wrong observation early on can steer every later step, and nothing in the basic pattern checks the final answer.
Predictable work
When the steps are known in advance, a fixed workflow does the same job with fewer calls and is easier to test.
How Motiontic's engineers apply it
Motiontic's engineers treat ReAct as the default shape of an agent loop rather than a feature to add. The work is in the parts the pattern leaves open: which tools the agent gets, how observations are trimmed before they reach the model, when the loop must stop, and what checks the answer.
Chain-of-thought prompting compared with a ReAct agent
| Chain-of-thought prompting | ReAct agent | |
|---|---|---|
| Uses external tools | No | Yes, between reasoning steps |
| Source of facts | The model's training data | Tool results it observes during the run |
| Model calls | One | One per step |
| Typical failure | Confident but invented facts | Loops or follows a bad observation |
| Fits | Reasoning over information already in the prompt | Tasks that need lookups or actions in other systems |
Common questions
What does ReAct stand for?
ReAct stands for reasoning and acting. It describes a model that alternates between writing a reasoning step and taking an action, then reads the result before continuing.
Is ReAct still used?
The pattern is everywhere, but the name less so. Most tool-calling agents run a reason, act and observe loop natively, and frameworks such as LangGraph have retired helpers named after ReAct in favour of a general agent constructor.
What is the difference between a ReAct agent and a tool-calling agent?
The original ReAct agent wrote its actions as text that code had to parse. A tool-calling agent returns actions as structured calls that match a schema. The loop is the same; the interface is sturdier.
Related
Agent loop
Model call, tool call, result, repeat: the cycle at the centre of every agent.
Tool calling
How a model asks software to do something: a named tool, typed arguments, a result.
Think tool
A tool that does nothing, so the model has somewhere to stop and think between actions.
LangChain vs LangGraph
The agent layer and the runtime underneath it, compared after the 1.0 releases.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Plan-and-execute agent
Plan the whole task first, then execute the steps, and replan only when something breaks.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.