ReAct (Reason+Act)

The defining structural characteristic of ReAct is its interleaved sequence of Thought, Action, and Observation steps. Unlike pure reasoning (Chain-of-Thought) or pure acting, ReAct prompts the model to generate a reasoning trace before deciding on an action, then feeds the result of that action back into the context window as an observation to inform the next reasoning step. This creates a dynamic feedback loop where the model can course-correct based on real-world data. For example, a ReAct agent might think 'I need to find the CEO of Apple,' act by calling a search tool, observe the Wikipedia result, then reason again to verify the information before answering.

ReAct requires a predefined action space—a discrete set of tool calls the model is allowed to invoke. These actions are typically expressed as structured commands within the prompt, such as:

• search[query]: Retrieves information from an external knowledge base
• lookup[string]: Finds the next term in a sequence
• finish[answer]: Signals task completion and returns the final answer The model must learn to map its internal reasoning to the correct action from this space. This constraint prevents unbounded behavior and ensures the agent interacts with external systems in a predictable, parseable format.

A primary benefit of ReAct is its ability to ground reasoning in retrieved facts rather than relying solely on parametric knowledge. When the model encounters uncertainty, it can issue a search action to query an external knowledge base like Wikipedia. This retrieved context is then explicitly incorporated into the subsequent reasoning trace. Research by Yao et al. (2022) demonstrated that ReAct significantly reduces hallucination rates on knowledge-intensive tasks like HotpotQA compared to pure reasoning baselines, because the model is forced to cite and reconcile external evidence before finalizing its answer.

ReAct demonstrates that reasoning and acting are mutually reinforcing. Reasoning traces help the model decompose complex goals, track progress, and handle exceptions. Actions provide the model with new information that updates its internal state. This synergy is particularly powerful for:

• Multi-hop question answering: Reasoning decomposes the question, actions retrieve intermediate facts
• Interactive decision-making: Reasoning plans the next move, actions execute it in an environment
• Fact verification: Reasoning identifies what needs checking, actions retrieve evidence Without reasoning, actions become blind trial-and-error. Without actions, reasoning becomes ungrounded speculation.

ReAct is implemented through carefully engineered prompts that provide few-shot examples of the Thought-Action-Observation pattern. A typical prompt includes:

• A system instruction defining the available actions
• 2-4 human-written trajectories showing correct interleaving
• A final user query that the model must complete The few-shot examples are critical—they teach the model when to reason versus when to act, and how to format its outputs. Poorly constructed examples can lead to reasoning-action misalignment, where the model either overthinks without acting or acts without sufficient reasoning.

ReAct occupies a distinct position in the prompting landscape:

• Chain-of-Thought (CoT): Reasoning only, no external actions. Prone to hallucination and fact-confabulation on knowledge-intensive tasks.
• Act-Only: Actions only, no explicit reasoning. Fails on tasks requiring multi-step planning or error recovery.
• ReAct: Combines both. Achieves the best of both worlds—structured problem decomposition from CoT and factual grounding from tool use. Empirical results show ReAct outperforms both baselines on tasks requiring dynamic information seeking, though pure CoT can be superior when internal knowledge is sufficient and tool calls introduce latency.

A complementary prompting technique where a language model drafts a response, then generates independent fact-checking questions to systematically verify its own output. Unlike ReAct's interleaved reasoning, CoVe operates as a post-hoc self-audit loop, decomposing claims into verifiable units and cross-referencing them against internal knowledge or retrieved evidence. This creates a two-stage grounding pipeline: ReAct handles real-time tool interaction during generation, while CoVe validates the final output's factual integrity.

The foundational architecture that grounds language model outputs in external knowledge bases by retrieving relevant documents before generation. ReAct extends RAG by adding a decision-making loop: instead of a single retrieval step, ReAct agents can perform multi-hop retrieval, evaluate intermediate results, and decide whether to search again, use a different tool, or finalize the answer. This transforms RAG from a static augmentation into a dynamic, self-directed research process.

A metric evaluating how precisely a generated statement's citations point to the specific source passages that directly support it. In ReAct systems, attribution fidelity becomes critical because the agent's observable reasoning trace creates an auditable chain from conclusion back to source. High attribution fidelity means each action's output can be deterministically traced to the evidence that triggered it, enabling enterprise-grade compliance and auditability for autonomous decisions.

A measure of uncertainty in a language model's output that clusters token-level predictions by semantic meaning before calculating entropy. This distinguishes between lexical variation (different words, same meaning) and genuine factual indecision (different meanings). In ReAct architectures, semantic entropy serves as a trigger mechanism for tool invocation: when entropy exceeds a threshold, the agent recognizes its own uncertainty and initiates an action to gather external grounding evidence rather than confabulating.

An open standard developed by Anthropic that defines a universal interface for connecting AI models to external tools and data sources. MCP provides the standardized action layer that ReAct reasoning traces can call: instead of custom API integrations for each tool, ReAct agents use MCP's client-server architecture to discover available tools, understand their schemas, and execute calls securely. This decouples the reasoning loop from tool implementation, enabling interoperable grounding across different systems.

A training methodology where a language model self-critiques and revises outputs based on a predefined set of principles. When combined with ReAct, Constitutional AI provides the normative guardrails for the reasoning trace: the agent not only reasons about factual correctness but also evaluates its planned actions against safety and ethical constraints before execution. This creates a dual-layer oversight mechanism—ReAct handles factual grounding, while Constitutional principles govern behavioral alignment.