What is the ReAct framework?

ReAct (Reasoning + Acting) interleaves chain-of-thought reasoning with tool calls. The model emits a Thought, an Action that invokes a tool, then receives an Observation, repeating until it has enough information to answer. It reduces hallucination by grounding reasoning in real tool results.

What is the Thought/Action/Observation loop?

Each cycle has three parts. Thought is the model's private reasoning, Action is a structured tool call, and Observation is the tool's returned result fed back in. The loop continues until the model emits a final answer instead of an action.

Why do I need an iteration cap?

Without a maximum, an agent can loop indefinitely on an unsolvable task, burning tokens and money. The cap forces the model to either answer or report failure within a bounded number of steps.

Does this run the agent for me?

No. It only writes the prompt. You still wire the Thought/Action/Observation loop in your own code and execute the tool calls the model requests.

ReAct Framework Prompt Builder

ReAct framework prompt builder

ReAct — short for Reasoning and Acting — is the dominant pattern for tool-using LLM agents. Rather than answering in one shot, the model alternates between thinking out loud and calling tools, grounding each step in real results. This builder assembles the system prompt that defines that loop: the turn format, the tools the agent may call, an iteration ceiling, and the signal that ends the loop with a final answer.

How it works

The generated prompt declares a strict turn structure: Thought (private reasoning), Action (a tool call written as tool_name[input]), and Observation (the result you feed back). The model repeats this cycle until it has enough to respond, at which point it emits Final Answer: in the format you chose. Your tool list is rendered as a name-and-description block so the model knows exactly what is callable and what each returns. The max iterations value is stated explicitly so the model knows to converge or report that it could not finish.

Tips and notes

Describe tool outputs, not just inputs. “search[query] → returns top 5 web snippets” tells the model how to use the Observation, which sharpens its next Thought.
Keep the tool set small. Three or four well-scoped tools outperform a dozen overlapping ones; the model picks more reliably from a short menu.
Match the loop in code. The prompt expects your runtime to detect an Action, run the tool, and inject the Observation. The cap and format only work if your harness honors them.
Use the final-answer format to parse results. A fixed Final Answer: prefix makes it trivial to extract the response programmatically.