Python: Native support for Bedrock-hosted models (Anthropic, Cohere, etc.) in Python SDK

[duttastunil]

The current Python SDK for the Microsoft Agent Framework requires users to route requests to Bedrock models via the generic OpenAI-compatible API interface. This approach has several drawbacks:

1. It does not natively recognize specific model IDs (e.g., anthropic.claude-3-5-sonnet-20240620-v1:0).
2. It may lack native support for provider-specific parameters or functionality (like Anthropic's specific messaging API structure, context windows, or token cost tracking).
3. It creates a less intuitive configuration experience compared to native integrations for Azure OpenAI or generic OpenAI services.

I would like to request dedicated, native client classes and configuration support within the Python SDK to directly integrate models hosted on Amazon Bedrock.

This should allow users to easily select and configure specific models (Anthropic Claude variants, Amazon Titan, Cohere models, etc.) using their native Bedrock model IDs/ARN and AWS credentials, without relying on the generic OpenAI compatibility layer.

This solution should include:

1. A dedicated BedrockClient or similar abstraction within the framework.
2. The ability to easily pass AWS authentication details (IAM roles/keys) to connect to the Bedrock runtime client (boto3).

The goal is to provide a first-class, seamless experience for developers building agents on AWS infrastructure using the Microsoft Agent Framework, leveraging the full capabilities of Bedrock-hosted models.

[sciutand]

FYI.. for dotnet (only mentioning because this issue also has the .NET label) AWS provides native support for Microsoft Extension Ai with this package.
https://www.nuget.org/packages/AWSSDK.Extensions.Bedrock.MEAI/4.0.3.1

A major limitation though is that it does not currently support response format. (because bedrock api doesn't apparently?)

[eavanvalkenburg]

For anthropic on Bedrock you can already use the AnthropicClient by passing in the AsyncAnthropicBedrock instead of a generic AsyncAnthropic. We are working on making that simpler, and also on other connectors, so thanks for raising this issue, we will use it to track.

[duttastunil]

@eavanvalkenburg - Thank you for providing the details. I was also implementing the BedrockChatClient which can support multiple models on bedrock, So that when we create the the agent, we just need to pass the BedrockChatClient() as below

The usage code is available here - https://github.com/duttastunil/agent-framework/blob/2524/python/samples/bedrock_chatclient/main.py

The BedrockChatClient code is available here - https://github.com/duttastunil/agent-framework/blob/2524/python/packages/bedrock/agent_framework_bedrock/_chat_client.py

Please let me know if you want to utilize this implementation.

[duttastunil]

The code comparison is available below.

https://github.com/microsoft/agent-framework/compare/main...duttastunil:agent-framework:2524?expand=1

[eavanvalkenburg]

Yeah, @duttastunil go ahead and create the PR, we'll review, but I had a quick look and it all looks good, following the conventions we have. The only thing is that we would like these clients to be available in agent_framework.amazon in case we add other amazon assets (for instance vector stores, that might be outside bedrock). I'm okay with the package name being agent-framework-bedrock