ReAct prompting pattern
ReAct, short for Reasoning and Acting, is an agent pattern in which a language model alternates between planning or reasoning, taking an allowed action, and incorporating the resulting observation before deciding what to do next. The loop gives the model a way to gather external information and revise a plan rather than relying on a single response from its internal knowledge.
Origin and context
The original paper proposed a shared trajectory of reasoning traces and actions. It evaluated the method on question answering and fact verification, where actions retrieved information, and on interactive tasks in ALFWorld and WebShop. The paper reported that reasoning helped direct actions while observations helped ground later reasoning. ReAct subsequently became a common reference architecture in agent frameworks and tutorials.
Why it matters
ReAct makes interaction part of solving a task: the model can seek a missing fact, observe the result, and change its next action. That differs from writing a plan once and executing it unchanged. Its value is the feedback between steps, not a promise that longer reasoning or more tool calls will improve every answer.
Example
As an illustrative question-answering workflow, an agent searching for a person's birthplace can first retrieve a biography, notice that it identifies a region but not a town, and make a more specific search. The next observation can resolve the gap or reveal conflicting information. This example illustrates the retrieval-and-revision pattern; it is not a reported benchmark result.
How it differs
Tool Use and Function Calling
Tool or function calling is the interface through which a model requests an external operation. ReAct is a broader iterative control pattern that can use such calls repeatedly, observe their results, and revise the next step. A single function call is not automatically a ReAct loop.
Maturity and evidence
Maturity is rated 3. ReAct has a clear peer-reviewed origin and is implemented across multiple agent ecosystems, but the label covers implementations with materially different prompts, state handling, stopping rules, and tool interfaces. The pattern is established; its operational behavior is not standardized.
Limits and open questions
A feedback loop can still follow an incorrect interpretation or fail to recover from an unhelpful observation. The original experiments show task-dependent strengths and weaknesses, including sensitivity to the information retrieved. ReAct names an interaction pattern, not a security boundary or a guarantee of correct execution. Assess the resulting actions and answers on the intended task, rather than treating a plausible-looking trajectory as proof of success.
Related terms
References
- ReAct: Synergizing Reasoning and Acting in Language ModelsarXiv · 2022-10-06 · class A
- What is a ReAct Agent?IBM · 2025 · class B
Last updated: 2026-09-05
This term is also covered in the Skills Atlas as agentic planning task decomposition skill.
This term is also covered in the Skills Atlas as llm function calling skill.
This term is also covered in the Skills Atlas as agent evaluation skill.