AWS | Community | How to use Reasoning with Claude 3.7 Sonnet on Amazon Bedrock

Also available in: C# | Java | Python

Have you ever wondered what goes on inside Claude when answering a complex question? With Claude 3.7 Sonnet's new reasoning feature, you can now peek behind the curtain and see the AI's step-by-step thought process. This capability offers unprecedented transparency into how Claude approaches problems, breaks them down, and reaches conclusions.

In this guide, we'll build a JavaScript application that taps into this powerful capability through Amazon Bedrock. You'll learn how to enable reasoning, allocate appropriate computational resources for it, and process the structured responses that include both Claude's internal reasoning and its final answer. Whether you're debugging prompts, verifying calculations, or using Claude as an educational tool, understanding its reasoning process can significantly enhance your AI interactions.

👨‍💻 Want to jump straight to the code? If you're familiar with Claude and Amazon Bedrock and just need the technical details, skip directly to the step-by-step implementation guide.

What is Reasoning?

Reasoning is Claude's ability to explicitly work through complex problems step by step before providing an answer. When enabled, Claude doesn't just give you the final response – it reveals its complete thought process, similar to how a human might think aloud when solving a problem.

This feature allows Claude to:

Break down complex questions into manageable components
Explore multiple approaches to solving a problem
Analyze information systematically and draw logical conclusions
Show its work, particularly valuable for mathematical or logical problems
Identify potential flaws in its own thinking

Reasoning is particularly useful for problems that require multi-step thinking, such as complex mathematical calculations, logical puzzles, or nuanced analytical questions. By exposing Claude's internal thought process, you gain transparency into how the AI arrives at its conclusions, which can help build trust and provide educational value.

When reasoning is enabled, Claude allocates a portion of its computational resources (specified by your token budget) to this explicit thinking process, which is then returned to you alongside the final answer. This gives you unprecedented visibility into the AI's problem-solving approach.

Prerequisites

Before we begin, make sure you have:

An AWS account with access to Amazon Bedrock
Node.js installed
The AWS CLI installed and configured
Access to Claude 3.7 Sonnet enabledin the region you want to use¹

¹ At the time of writing, Claude 3.7 is available in us-east-1, us-east-2, and us-west-2.

Setting Up Dependencies

You'll need to install the AWS SDK for JavaScript v3. Here's how to include the required dependencies:

Or add the following to your package.json file:

Step 1: Create a Basic Project Structure

Create a new JavaScript file (e.g., claude_reasoning.js) where we'll implement our code to interact with Claude's reasoning capability:

Step 2: Set Up the Amazon Bedrock Client

First, we need to create a Bedrock runtime client that will handle our API requests:

Step 3: Create a Message for Claude

Next, we'll create a simple message to send to Claude:

Step 4: Configure Reasoning Parameters

Here's where the magic happens - we need to enable reasoning and set a token budget for it:

The budget_tokens parameter defines the maximum number of tokens Claude can use for its reasoning process. Adjust this value based on the complexity of your prompt.

Step 5: Send the Request to Claude

Now we'll send our prompt along with the reasoning configuration:

The additionalModelRequestFields parameter is what allows us to pass the reasoning configuration. Notice that we're using the ConverseCommand class from the AWS SDK for JavaScript v3.

Step 6: Process the Response

Claude will return both its reasoning and final response, which we need to extract:

The response contains multiple content blocks, with the reasoning in one block and the final response text in another.

Step 7: Display the Results

Finally, we'll display both the reasoning process and the final answer:

Complete Example

Here's the complete working example

Note: Make sure to run this code with a recent version of Node.js that supports ES modules. You may need to add "type": "module" to your package.json file or use the .mjs extension for your JavaScript file.

Example Output

When you run this code, you'll see output similar to:

Summary and next steps

The reasoning capability transforms how we interact with AI by providing valuable insight into the model's thinking process. By following the steps in this guide, you can now access and leverage this powerful feature through Amazon Bedrock using JavaScript.

The reasoning feature opens up exciting possibilities to enhance trust, improve outcomes, and gain deeper insights from your AI interactions. For more advanced use cases, consider:

Adjusting the token budget based on your problem complexity – more complex problems may benefit from larger reasoning budgets
Using the reasoning output to validate multi-step calculations or complex analytical processes
Comparing different reasoning approaches by adjusting your prompts
Integrating reasoning with other Claude capabilities like function calling for powerful, transparent AI solutions
Using reasoning as an educational tool to understand expert-level problem-solving approaches

By incorporating reasoning into your applications, you're not just getting answers – you're gaining insight into the full problem-solving journey. This transparency can help build trust with users and provide richer, more valuable AI interactions.

Happy coding!

Any opinions in this post are those of the individual author and may not reflect the opinions of AWS.

Select your cookie preferences

Site Terms, Privacy, and more.

How to use Reasoning with Claude 3.7 Sonnet on Amazon Bedrock - JavaScript Edition

Learn how to use Claude 3.7 Sonnet's step-by-step thinking process with the AWS SDK for JavaScript v3. This tutorial walks you through a practical example showing how to receive and process Claude's internal thought process in your own applications.