ReAct
An agent pattern that interleaves reasoning and acting: the model thinks, takes a tool action, observes the result, and reasons again — looping to a goal.
ReAct is the loop the agentic web is built to serve: every markdown twin, JSON API and callable tool is an 'action' an agent takes between reasoning steps.
- term
- ReAct
- category
- core
- short_def
- An agent pattern that interleaves reasoning and acting: the model thinks, takes a tool action, observes the result, and reasons again — looping to a goal.
- long_def
- ReAct (Reason + Act) combines chain-of-thought with tool use so an agent alternates 'thought' and 'action' steps: it reasons about what to do, calls a tool, reads the observation, and updates its plan. It is the canonical loop most tool-using agents implement.
- see_also
agentic-looptool-usechain-of-thought- etymology_origin
- — verify-against-primary-at-build ↗ https://arxiv.org/abs/2210.03629 — 'ReAct' introduced by Yao et al., 2022 (Princeton / Google)
- related_to
agentic-looptool-usechain-of-thoughtai-agent- contrast_with
- Unlike plain chain-of-thought, which only reasons, ReAct also acts between reasoning steps — grounding each thought in a real observation from a tool.
- example
- A ReAct agent answering a live question reasons ('I need current data'), calls a search tool, reads the result, then reasons again before answering.
- source
- https://arxiv.org/abs/2210.03629
- status
- active
- why_it_matters
- ReAct is the loop the agentic web is built to serve: every markdown twin, JSON API and callable tool is an 'action' an agent takes between reasoning steps.
- sameAs
https://arxiv.org/abs/2210.03629- bridge_entity
- agentic-loop
- last_verified
- 2026-07-06
- md_twin
- /glossary/react-pattern.md
last verified · by Özden Erdinc