---
title: "ReAct Agent Loop"
type: "prompt"
provider: "community"
category: "agent-harnesses"
source: "https://github.com/topics/react-agent"
license: "MIT"
tags: ["agent", "react", "tools", "reasoning"]
dateAdded: "2026-06-13"
featured: true
origin: "community"
language: "en"
description: "The classic Reason+Act harness: interleave thoughts, tool calls and observations until the task is solved. A reusable scaffold for tool-using agents."
---

A minimal harness that drives a tool-using agent with the ReAct pattern.

## Loop

Repeat until you can answer, then stop:

```
Thought: <reason about what to do next>
Action: <tool_name>[<input>]
Observation: <result returned by the tool>
```

When you have enough information:

```
Thought: I can now answer.
Answer: <final answer>
```

## Rules

- Take exactly one Action per step, then wait for the Observation before continuing.
- Never fabricate Observations — only the tool runtime produces them.
- If a tool errors, reason about why in the next Thought and adapt.
- Keep Thoughts short; they are scratchpad, not the final answer.

**Available tools:** {{list each tool with its name, input format and what it returns}}