ReAct Agent

ReAct Agent Definition: What It Actually Means

In short

A ReAct agent is an LLM-based agent that alternates between generating natural language reasoning traces and executing discrete actions — such as search queries or API calls — using the result of each action to inform the next reasoning step. The term ReAct is a portmanteau of Reasoning and Acting, introduced by Yao et al. in 2022.

A ReAct agent is an AI agent architecture that interleaves natural language reasoning with discrete tool-use actions in a repeating loop. The loop continues — Thought, then Action, then Observation — until the model produces a Final Answer.

The name ReAct is a portmanteau of Reasoning and Acting. It has no connection to the React JavaScript framework.

The architecture was introduced in the paper "ReAct: Synergizing Reasoning and Acting in Language Models" by Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. The paper was submitted to arXiv in October 2022 (arXiv:2210.03629) and presented at ICLR 2023.

The loop has three components, each with a distinct role:

• Thought — The model writes out its internal reasoning in natural language, deciding what to do next and why.
• Action — The model calls a tool (e.g., a Wikipedia search, calculator, database query, or REST API).
• Observation — The result of the tool call is returned and appended to the model's context window as an observation.

This loop repeats until the model decides it has sufficient information to emit a Final Answer. The number of iterations is not fixed — it runs until task completion or a maximum step limit is reached.

The key innovation over prior work is the interleaving: reasoning informs which action to take, and action results correct and update the reasoning. This grounds the model in real-world data rather than relying purely on parametric (trained) knowledge — the root cause of most LLM hallucinations on factual tasks.

The original paper benchmarked ReAct on HotpotQA (multi-hop question answering), FEVER (fact verification), ALFWorld (interactive household tasks), and WebShop (online shopping simulation).

How ReAct Agents Work: Architecture and Tools

In short

A ReAct agent is built on three components: a structured prompt template that encodes the Thought-Action-Observation format, an LLM that generates reasoning and action strings, and a tool executor that runs actions and returns observations. Modern frameworks like LangGraph provide ready-made ReAct executors.

In production, a ReAct agent consists of three cooperating components working in sequence on every loop iteration.

• Structured prompt — Includes the task, a list of available tools with descriptions, and few-shot TAO traces to calibrate output format.
• LLM backbone — GPT-4, Gemini, Claude, or similar — generates the next Thought and Action token sequence based on the full context window.
• Tool dispatcher/executor — Parses the Action string, routes it to the correct tool, executes it, and appends the Observation to the context window.

Tools are described to the model in the system prompt using a structured format: name, description, and input schema. The model learns correct invocation from the few-shot examples and tool descriptions — making description quality a direct driver of agent accuracy.

A practical constraint: the context window fills with each TAO iteration. On long-horizon tasks with many tool calls, this can exhaust the model's context limit — a key architectural consideration when designing agents for complex workflows.

Common Tools Used in ReAct Agents

ReAct vs. Other Agent Architectures: Key Differences

In short

ReAct differs from chain-of-thought prompting by adding grounded tool use, from function-calling agents by making reasoning explicit and auditable, and from plan-and-execute agents by selecting each action dynamically rather than committing to a plan upfront.

ReAct occupies a specific position in the agent architecture landscape. Understanding how it compares to alternatives is essential for choosing the right pattern for a given task.

ReAct vs. Other Agent Architectures

The most important comparison for practitioners is ReAct vs. function calling — they are often confused because both involve tool use. The critical difference is that ReAct exposes the model's reasoning in the Thought trace, while function calling keeps it opaque inside the model weights.

For enterprise deployments in regulated industries — finance, healthcare, legal — the auditability of Thought traces is not just a nice-to-have. It is frequently a compliance requirement. For a broader view of where ReAct fits in the agentic landscape, see our overview of what is agentic AI and our comparison of AI agent architecture patterns.

When to Use a ReAct Agent — and When Not To

In short

ReAct is the right architecture for multi-step tasks requiring external data retrieval, auditable reasoning, or adaptive decision-making. It is not optimal for single-step tasks, latency-sensitive applications, or tasks fully solvable with parametric knowledge alone.

ReAct is a powerful architecture, but it is not always the right choice. Selecting the appropriate agent pattern is one of the most consequential decisions in an enterprise AI deployment.

Use ReAct when:

• The task requires multiple external data sources — e.g., cross-referencing supplier data, market data, and internal documents simultaneously.
• You need auditability — the Thought traces provide an interpretable record of why each action was taken, critical for regulated industries.
• The task is open-ended or uncertain — where the number and type of steps cannot be predetermined before execution begins.
• You are building research, analysis, or due diligence workflows where the agent must synthesise information across many sources.

Do not use ReAct when:

• The task is single-step and can be answered with one tool call — function calling is faster and simpler.
• You have strict latency requirements — each TAO iteration adds a round-trip to the LLM, meaning ReAct is inherently multi-call.
• The task is fully within the model's parametric knowledge — adding tools introduces unnecessary complexity and potential failure points.
• You are operating in a cost-constrained environment — multiple LLM calls per task multiply API costs, especially with frontier models.

In practice, most enterprise agent systems use ReAct as one component within a larger orchestration layer — not as a monolithic architecture for all tasks. See our guide on what is an AI agent for a broader framework for making these architectural decisions.

ReAct Agent Limitations and Security Risks

In short

ReAct agents face three production-critical limitations: context window exhaustion in long-horizon tasks, vulnerability to indirect prompt injection attacks (IBM Research, 2025), and error propagation when an early Observation is incorrect or misleading.

ReAct is powerful in the right context, but it introduces specific failure modes that enterprise teams must design for before going to production.

Limitation 1: Context Window Exhaustion

Every TAO iteration appends Thoughts, Actions, and Observations to the context window. On long-horizon tasks with many tool calls, this can exhaust the model's context limit — causing truncation, degraded reasoning, or task failure. Mitigation strategies include summarising earlier observations, limiting maximum iterations, or using models with larger context windows.

Limitation 2: Error Propagation

If an early Observation returns incorrect or misleading data — due to a bad search result, a stale API response, or a misformatted tool output — the model's subsequent Thoughts will be built on a faulty foundation. Unlike humans, ReAct agents do not inherently detect when a source is unreliable.

Limitation 3: Indirect Prompt Injection (Critical)

IBM Research (2025) demonstrated that ReAct agents are vulnerable to indirect prompt injection attacks via a "foot-in-the-door" technique. An attacker can embed adversarial instructions inside a web page or document that the agent retrieves as an Observation — causing the agent to execute unintended actions as if they were legitimate instructions.

Additional production concerns include: infinite loops (if the agent never converges on a Final Answer), cost overruns from excessive tool calls, and non-determinism across runs. Comprehensive mitigation requires a production checklist approach — see our AI production deployment checklist and our dedicated guide to AI agent security risks.

ReAct Agents in Enterprise AI: 2025–2026 Landscape

In short

ReAct is one of the most commonly deployed agentic patterns in production as of 2026, with 57% of engineering teams running agents in production (LangChain, 2026) and the global AI agents market projected at $182.97B by 2033 (Grand View Research, 2026). Enterprise adoption is accelerating across research, customer service, data analysis, and procurement use cases.

The AI agents market is growing rapidly. Grand View Research projects the global AI agents market will reach $182.97 billion by 2033 — driven by enterprise adoption of agentic architectures across industries.

LangChain's 2026 State of Agent Engineering survey found that 57% of engineering teams now have AI agents in production — up substantially from prior years. ReAct remains one of the most commonly implemented patterns, largely because of its framework support maturity and interpretability advantages.

Enterprise use cases where ReAct agents are delivering production value in 2025–2026 include:

• Research and due diligence — Agents that search, retrieve, and synthesise information across multiple sources for investment analysis, procurement evaluation, or competitive intelligence.
• Data analysis workflows — Agents that query databases, run Python analysis, and produce reports — replacing manual analyst steps.
• Customer service escalation handling — Agents that look up order history, check policies, and draft responses with full context before human review.
• Procurement intelligence — Agents that cross-reference supplier databases, pricing APIs, and internal ERP data to support sourcing decisions.

For organisations evaluating where ReAct fits in their broader AI strategy, the key question is not whether to use ReAct — but which tasks genuinely require the Thought-Action-Observation loop versus simpler patterns. See our enterprise AI strategy framework for a structured approach to these architectural decisions, and our overview of best AI agent frameworks in 2026 for the implementation tooling landscape.