Prompt Engineering Patterns: ReAct Pattern — Reasoning and Acting with Tool Use for LLM Agents

ReAct, short for Reasoning and Acting, is a prompting pattern that interleaves chain-of-thought reasoning with tool actions. The model produces a Thought explaining what to do, an Action calling a tool, an Observation receiving the tool's result, then another Thought, and so on until it reaches a final answer. ReAct was introduced by Yao and colleagues in 2022.

ReAct lets the model reason about what information it needs, take an action to get it, observe the result, then reason about the next step. Reasoning targets what to retrieve, while retrieval grounds the reasoning in real data. This synergy outperforms both pure reasoning and pure tool use, especially on tasks like fact verification, multi-hop question answering, and web navigation.

A ReAct trace alternates Thought, Action, and Observation steps. For example: Thought: I need to find when the iPhone was first released. Action: search for iPhone release date. Observation: The first iPhone was released in 2007. Thought: I have the answer. Final Answer: 2007. The model produces Thoughts and Actions, while the system executes the action and provides the Observation.

ReAct beats pure CoT on knowledge-intensive tasks because it can retrieve fresh information instead of relying on training data. CoT reasoning amplifies the model's existing knowledge but cannot fix gaps or outdated facts. ReAct addresses this by letting the model fetch what it needs at runtime, dramatically reducing hallucinations on factual questions.

ReAct exposes the model's decision process as an explicit trace of thoughts and actions. Developers can inspect why the model chose a specific tool, what information it retrieved, and how it reasoned about results. This makes debugging much easier than with opaque single-shot answers and lets humans intervene to correct mistakes mid-trace.

ReAct is a prompting pattern that any model can implement through structured text. Function calling is a model API feature where the model produces structured tool invocations with validated arguments. Modern LLM APIs from OpenAI, Anthropic, and Google support function calling natively, which is more reliable than text-based ReAct parsing. Many production systems use function calling as the underlying transport for ReAct-style logic.

In production, implement ReAct by giving the model a list of available tools with descriptions, parsing its responses for action calls, executing those calls, feeding results back into the conversation, and repeating until the model produces a final answer. Add a maximum iteration limit to prevent infinite loops, error handling for failed tool calls, and observability to debug failures.

ReAct interleaves thinking and acting at every step, deciding the next action based on the previous observation. Plan-and-Execute first generates a complete multi-step plan, then executes each step in order. ReAct adapts better when results are unexpected, while Plan-and-Execute is more efficient when the path is predictable. Many production agents use Plan-and-Execute at the high level and ReAct within steps.

Common ReAct failures include infinite loops where the agent repeats the same action, hallucinated tool calls with invented arguments, getting stuck on impossible subtasks, losing track of the original goal, producing thoughts that ignore previous observations, and runaway costs from too many iterations. Mitigations include step limits, tool argument validation, explicit goal tracking, and cost budgets. ---