Skip to content
Multi-Agent Orchestrator

Bedrock Classifier

The Bedrock Classifier is the default classifier used in the Multi-Agent Orchestrator. It leverages Amazon Bedrock’s models through Converse API providing powerful and flexible classification capabilities.

Features

  • Utilizes Amazon Bedrock’s models through Converse API
  • Configurable model selection and inference parameters
  • Supports custom system prompts and variables
  • Handles conversation history for context-aware classification

Default Model

The classifier uses Claude 3.5 Sonnet as its default model:

BEDROCK_MODEL_ID_CLAUDE_3_5_SONNET = "anthropic.claude-3-5-sonnet-20240620-v1:0"

Model Support for Tool Choice

The BedrockClassifier’s toolChoice configuration for structured outputs is only available with specific models in Amazon Bedrock. As of January 2025, the following models support tool use:

  • Anthropic Models:

    • Claude 3 models (all variants except Haiku)
    • Claude 3.5 Sonnet (anthropic.claude-3-5-sonnet-20240620-v1:0)
    • Claude 3.5 Sonnet v2

  • AI21 Labs Models:

    • Jamba 1.5 Large
    • Jamba 1.5 Mini

  • Amazon Models:

    • Nova Pro
    • Nova Lite
    • Nova Micro

  • Meta Models:

    • Llama 3.2 11b
    • Llama 3.2 90b

  • Mistral AI Models:

    • Mistral Large
    • Mistral Large 2 (24.07)
    • Mistral Small

  • Cohere Models:

    • Command R
    • Command R+

When using other models:

  • The tool configuration will still be included in the request
  • The model won’t be explicitly directed to use the analyzePrompt tool
  • Response formats may be less consistent

For the most up-to-date list of supported models and their features, please refer to the Amazon Bedrock Converse API documentation.

Python Package

If you haven’t already installed the AWS-related dependencies, make sure to install them:

Terminal window
pip install "multi-agent-orchestrator[aws]"

Basic Usage

By default, the Multi-Agent Orchestrator uses the Bedrock Classifier:

import { MultiAgentOrchestrator } from "multi-agent-orchestrator";


const orchestrator = new MultiAgentOrchestrator();

System Prompt and Variables

Full Default System Prompt

The default system prompt used by the classifier is comprehensive and includes examples of both simple and complex interactions:

You are AgentMatcher, an intelligent assistant designed to analyze user queries and match them with
the most suitable agent or department. Your task is to understand the user's request,
identify key entities and intents, and determine which agent or department would be best equipped
to handle the query.


Important: The user's input may be a follow-up response to a previous interaction.
The conversation history, including the name of the previously selected agent, is provided.
If the user's input appears to be a continuation of the previous conversation
(e.g., "yes", "ok", "I want to know more", "1"), select the same agent as before.


Analyze the user's input and categorize it into one of the following agent types:
<agents>
{{AGENT_DESCRIPTIONS}}
</agents>
If you are unable to select an agent put "unknown"


Guidelines for classification:


    Agent Type: Choose the most appropriate agent type based on the nature of the query.
    For follow-up responses, use the same agent type as the previous interaction.
    Priority: Assign based on urgency and impact.
        High: Issues affecting service, billing problems, or urgent technical issues
        Medium: Non-urgent product inquiries, sales questions
        Low: General information requests, feedback
    Key Entities: Extract important nouns, product names, or specific issues mentioned.
    For follow-up responses, include relevant entities from the previous interaction if applicable.
    For follow-ups, relate the intent to the ongoing conversation.
    Confidence: Indicate how confident you are in the classification.
        High: Clear, straightforward requests or clear follow-ups
        Medium: Requests with some ambiguity but likely classification
        Low: Vague or multi-faceted requests that could fit multiple categories
    Is Followup: Indicate whether the input is a follow-up to a previous interaction.


Handle variations in user input, including different phrasings, synonyms,
and potential spelling errors.
For short responses like "yes", "ok", "I want to know more", or numerical answers,
treat them as follow-ups and maintain the previous agent selection.


Here is the conversation history that you need to take into account before answering:
<history>
{{HISTORY}}
</history>


Skip any preamble and provide only the response in the specified format.

Example Variable Replacements

AGENT_DESCRIPTIONS Example

aws-services-agent:Expert in AWS services configuration, troubleshooting, and best practices
security-agent:Handles security-related inquiries, compliance questions, and access management
billing-agent:Manages AWS billing inquiries, cost optimization, and pricing details
developer-support:Assists with code, SDK issues, and AWS service implementation

Extended HISTORY Examples

The conversation history is formatted to include agent names in the responses, allowing the classifier to track which agent handled each interaction. Each assistant response is prefixed with [agent-name] in the history, making it clear who provided each response:

Example 1: AWS Services to Billing Transition

user: How do I set up an S3 bucket with versioning?
assistant: [aws-services-agent] I'll help you set up S3 versioning. First, do you already have an S3 bucket created?
user: No, but how much will it cost?
assistant: [billing-agent] I can help explain S3 pricing. The cost depends on storage usage, versioning, and access patterns. What's your expected storage volume?
user: Around 500GB of files that change frequently

In this example, you can see how the history shows the transition from aws-services-agent to billing-agent, with each response clearly marked.

Example 2: Multiple Department Interaction

user: I can't access my EC2 instance
assistant: [aws-services-agent] Let's troubleshoot your EC2 access. Can you see the instance in your console?
user: Yes, but SSH keeps saying permission denied
assistant: [security-agent] This sounds like a security group or key pair issue. When was the last time you successfully connected to this instance?
user: Yesterday, but we rotated some keys
assistant: [security-agent] I'll help you verify your key configuration and permissions. Which key pair are you currently using to connect?

Here, the history shows the conversation moving between aws-services-agent and security-agent as the topic shifts from general access to security configuration.

The agent prefixing (e.g., [agent-name]) is automatically handled by the Multi-Agent Orchestrator when formatting the conversation history. This helps the classifier understand:

  • Which agent handled each part of the conversation
  • The context of previous interactions
  • When agent transitions occurred
  • How to maintain continuity for follow-up responses

Tool-Based Response Structure

Like the Anthropic Classifier, the BedrockClassifier uses a tool specification to enforce structured output from the model. This is a design pattern that ensures consistent and properly formatted responses.

The Tool Specification

{
  "toolSpec": {
    "name": "analyzePrompt",
    "description": "Analyze the user input and provide structured output",
    "inputSchema": {
      "json": {
        "type": "object",
        "properties": {
          "userinput": {"type": "string"},
          "selected_agent": {"type": "string"},
          "confidence": {"type": "number"}
        },
        "required": ["userinput", "selected_agent", "confidence"]
      }
    }
  }
}

Why Use Tools?

  1. Structured Output: Instead of free-form text, the model must provide exactly the data structure we need.
  2. Guaranteed Format: The tool schema ensures we always get:
    • A valid agent identifier
    • A properly formatted confidence score
    • All required fields
  3. Implementation Note: The tool isn’t actually executed - it’s a pattern to force the model to structure its response in a specific way that maps directly to our ClassifierResult type.

Example Response:

{
  "userinput": "How do I configure VPC endpoints?",
  "selected_agent": "aws-services-agent",
  "confidence": 0.95
}

Custom Configuration

You can customize the BedrockClassifier by creating an instance with specific options:

import { BedrockClassifier, MultiAgentOrchestrator } from "multi-agent-orchestrator";


const customBedrockClassifier = new BedrockClassifier({
  modelId: 'anthropic.claude-3-sonnet-20240229-v1:0',
  region: 'us-west-2',
  inferenceConfig: {
    maxTokens: 500,
    temperature: 0.7,
    topP: 0.9
  }
});


const orchestrator = new MultiAgentOrchestrator({ classifier: customBedrockClassifier });

The BedrockClassifier accepts the following configuration options:

  • model_id (optional): The ID of the Bedrock model to use. Defaults to Claude 3.5 Sonnet.
  • region (optional): The AWS region to use. If not provided, it will use the REGION environment variable.
  • inference_config (optional): A dictionary containing inference configuration parameters:
    • maxTokens (optional): The maximum number of tokens to generate.
    • temperature (optional): Controls randomness in output generation.
    • topP (optional): Controls diversity of output generation.
    • stopSequences (optional): A list of sequences that will stop generation.

Best Practices

  1. AWS Configuration: Ensure proper AWS credentials and Bedrock access are configured.
  2. Model Selection: Choose appropriate models based on your use case requirements.
  3. Region Selection: Consider using the region closest to your application for optimal latency.
  4. Inference Configuration: Experiment with different parameters to optimize classification accuracy.
  5. System Prompt: Consider customizing the system prompt for your specific use case, while maintaining the core classification structure.

Limitations

  • Requires an active AWS account with access to Amazon Bedrock
  • Classification quality depends on the chosen model and the quality of agent descriptions
  • Subject to Amazon Bedrock service quotas and pricing

For more information, see the Classifier Overview and Agents documentation.