Historical source material recovered on 2026-09-09. Capture dates and source limitations apply; recovery is not fresh verification.

ReAct Loop (Reason + Act)

The reasoning-action loop that sits at the heart of every agentic AI system. The LLM alternates between reasoning steps (thinking about what to do next) and action steps (calling a tool, observing the result). The most widely adopted architecture for agentic AI in 2026 ^[src-003].

The six-step cycle ^[src-003]

1. Receive goal — the agent receives a high-level objective from the user

2. Reason — the LLM analyses the goal and generates a plan or next action

3. Act — the agent calls a tool (web search, write code, query a database, call an API)

4. Observe — the agent reads the tool's output and updates its understanding

5. Iterate — steps 2–4 repeat until the goal is achieved or a stop condition is met

6. Return result — the agent delivers the final output to the user

Why it works

The ReAct loop gives an LLM two capabilities it can't have in a single prompt-response cycle ^[src-003]:

• Adaptation to unexpected results — the agent can react to what actually comes back from a tool call, not just what it predicted would happen
• Error recovery — a failed step can be re-reasoned and retried rather than blocking the whole task
• Long-horizon tasks — objectives that need dozens of sequential decisions become tractable because state is preserved across iterations

Where ReAct appears

• Single-agent systems: Claude Code runs a ReAct-style loop internally every time it's given a task — reason, call a tool (Read/Write/Bash), observe, continue.
• Multi-agent systems: each agent in Paperclip runs its own ReAct loop on top of its instruction files during a heartbeat. The orchestration layer coordinates between agents but each individual agent is still ReAct under the hood.
• Frameworks: LangGraph, AutoGen, OpenAI Agents SDK all implement variants of this loop as their primary execution model.
• Agentic vision: Gemini Robotics-ER 1.6 applies a ReAct-like structure to visual perception, zooming into images, pointing to relevant features, using code execution, and then interpreting the result ^[src-039].
• Physical agents: embodied robotics uses the same reason-act-observe pattern, but the observations are camera views and physical state, and the actions may be tool calls, VLA calls, or robot plan updates ^[src-039].

What ReAct is NOT ^[src-003]

• Not a prompt template — ReAct describes a runtime loop, not a single prompt pattern
• Not specific to any model — works with any LLM that can follow multi-step reasoning and call tools
• Not foolproof — it is vulnerable to compounded hallucinations when an early step introduces an error that propagates

Related entities

• Claude Code
• Paperclip

Related concepts

• Agentic AI
• Agent Orchestration
• Heartbeats
• Model Context Protocol
• Agentic Vision
• Embodied Reasoning
• Robotic Success Detection

Source references

• ^[src-003] Robin Cartier — "What is Agentic AI? A Complete Guide" (2026-03-10)
• ^[src-039] Laura Graesser and Peng Xu — "Gemini Robotics-ER 1.6: Powering real-world robotics tasks through enhanced embodied reasoning" (2026-04-14)

Historical recovered snapshot — 2026-08-03

Preserved from pr-123. Documentary recovery on 2026-09-09; claims below retain their historical source scope and were not freshly verified. Current synthesis above remains authoritative.

The original ReAct paper defines the loop more precisely as interleaved thought, action and observation trajectories. Reasoning maintains and revises plans, while environment actions retrieve evidence that grounds subsequent reasoning ^[src-263]. Its reported gains are benchmark-specific: 34-point and 10-point absolute success-rate improvements on ALFWorld and WebShop, while knowledge tasks benefited most from combining ReAct with chain-of-thought self-consistency ^[src-263].

This primary source corrects two overextensions: ReAct is a prompting and interaction paradigm as well as a runtime loop, and its inspectable traces improve diagnosis but do not replace permissions, stopping rules or output verification.