How to integrate AWS Bedrock Models like NOVA etc using Autogen ?

[amardeepjaiman]

Hi ,

I want to keep Autogen as common multi agent framework in my agentic application who can supports both Azure and AWS models ? I see that there is no dedicated AWSBedrockChatCompletionClient like AzureOpenAI, AzureAI etc.

is it not supported yet? any workaround?

Thanks,
Amardeep

[Lucas Kim (kimnamu)]

Thanks for laying out the Azure + AWS requirement so clearly, and for following up with the concrete Nova traces in #6479 and #6480. That made it easy to test the exact path you are on.

There is no generic Bedrock client in autogen-ext: Claude on Bedrock has AnthropicBedrockChatCompletionClient, and Nova (and the other Converse API models) go through SKChatCompletionAdapter on top of Semantic Kernel's BedrockChatCompletion. Plain chat works on main today; tool calling needs a small fix in KernelFunctionFromToolSchema, verified below.

What I ran (autogen main @ 027ecf0, semantic-kernel 1.36.0, boto3 1.40, us-east-1):

Route	Model	Result
SKChatCompletionAdapter + BedrockChatCompletion, create()	us.amazon.nova-lite-v1:0, nova-pro, nova-2-lite	200, 'Paris'
same, AssistantAgent with a tool, stock main	all three	args come back as {"city": {"name": "Paris"}}, pydantic validation error
same, with the patch below	us.amazon.nova-lite-v1:0	{"city": "Paris"}, The weather in Paris is 75 degrees.
same, reflect_on_tool_use=True	us.amazon.nova-lite-v1:0	Converse 400 The toolConfig field must be defined (the reflection call sends no tools)
OpenAIChatCompletionClient against bedrock-runtime.<region>.amazonaws.com/openai/v1	us.amazon.nova-lite-v1:0	model_not_found (that endpoint does not serve Nova)
AnthropicBedrockChatCompletionClient	us.anthropic.claude-haiku-4-5-20251001-v1:0	200 with anthropic<1; anthropic 1.6.0 raises TypeError ... 'temperature'

pip install "autogen-agentchat" "autogen-ext[semantic-kernel-aws]"

import asyncio

import boto3
from autogen_agentchat.agents import AssistantAgent
from autogen_core.models import ModelFamily
from autogen_ext.models.semantic_kernel import SKChatCompletionAdapter
from semantic_kernel import Kernel
from semantic_kernel.connectors.ai.bedrock import BedrockChatCompletion, BedrockChatPromptExecutionSettings


def get_weather(city: str) -> str:
    """Get the weather for a city."""
    return f"The weather in {city} is 75 degrees."


async def main() -> None:
    region = "us-east-1"  # boto3 uses its normal credential chain (profile, env vars, Bedrock API key)
    sk_client = BedrockChatCompletion(
        model_id="us.amazon.nova-lite-v1:0",
        runtime_client=boto3.client("bedrock-runtime", region_name=region),
        client=boto3.client("bedrock", region_name=region),
    )
    model_client = SKChatCompletionAdapter(
        sk_client,
        kernel=Kernel(),
        prompt_settings=BedrockChatPromptExecutionSettings(temperature=0.0, max_tokens=512),
        model_info={"function_calling": True, "json_output": False, "vision": True, "family": ModelFamily.UNKNOWN, "structured_output": False},
    )
    agent = AssistantAgent("assistant", model_client=model_client, tools=[get_weather])
    result = await agent.run(task="What is the weather in Paris?")
    print(result.messages[-1].content)


asyncio.run(main())

Why tool calls break on stock main: AssistantAgent hands the adapter ToolSchema dicts, so the adapter builds KernelFunctionFromToolSchema (added in #6637). That class reads required from properties instead of parameters, and gives InputVariable no type, so SK's build_from_type_name("") falls back to object. This is the toolConfig Bedrock actually received, before and after the patch:

before: "properties": {"city": {"type": "object", "description": "city"}}, "required": []
after:  "properties": {"city": {"type": "string", "description": "city"}}, "required": ["city"]

diff --git a/python/packages/autogen-ext/src/autogen_ext/tools/semantic_kernel/_kernel_function_from_tool.py b/python/packages/autogen-ext/src/autogen_ext/tools/semantic_kernel/_kernel_function_from_tool.py
index 1149195..5b8cb0e 100644
--- a/python/packages/autogen-ext/src/autogen_ext/tools/semantic_kernel/_kernel_function_from_tool.py
+++ b/python/packages/autogen-ext/src/autogen_ext/tools/semantic_kernel/_kernel_function_from_tool.py
@@ -70,17 +70,32 @@ async def tool_method(**kwargs: dict[str, Any]) -> Any:
         self._tool = tool
 
 
+# JSON Schema type names -> type names SK's KernelJsonSchemaBuilder.build_from_type_name understands.
+_JSON_TYPE_TO_SK_TYPE = {
+    "string": "str",
+    "integer": "int",
+    "number": "float",
+    "boolean": "bool",
+    "array": "list",
+    "object": "dict",
+}
+
+
 class KernelFunctionFromToolSchema(KernelFunctionFromPrompt):
     def __init__(self, tool_schema: ToolSchema, plugin_name: str | None = None):
-        properties = tool_schema.get("parameters", {}).get("properties", {})
-        required = properties.get("required", [])
+        parameters = tool_schema.get("parameters", {})
+        properties = parameters.get("properties", {})
+        required = parameters.get("required", [])
 
         prompt_template_config = PromptTemplateConfig(
             name=tool_schema.get("name", ""),
             description=tool_schema.get("description", ""),
             input_variables=[
                 InputVariable(
-                    name=prop_name, description=prop_info.get("description", ""), is_required=prop_name in required
+                    name=prop_name,
+                    description=prop_info.get("description", ""),
+                    is_required=prop_name in required,
+                    json_schema=_JSON_TYPE_TO_SK_TYPE.get(prop_info.get("type", "string"), "dict"),
                 )
                 for prop_name, prop_info in properties.items()
             ],

Raw outputs and notes

Adapter parameter metadata, offline (no network), before and after the patch:

before: city is_required=False schema_data={'type': 'object', 'description': 'city'}
after:  city is_required=True  schema_data={'type': 'string', 'description': 'city'}
        days is_required=False schema_data={'type': 'integer', 'description': 'days'}

Agent run on stock main with us.amazon.nova-lite-v1:0, then with the patch:

ToolCallRequestEvent -> [FunctionCall(id='tooluse_wxGDCKyZ1pEYH4ny25KzJ2', arguments='{"city": {"name": "Paris"}}', name='get_weather')]
ToolCallSummaryMessage -> 1 validation error for get_weatherargs ... Input should be a valid string

ToolCallRequestEvent -> [FunctionCall(id='tooluse_UclIncJCufeWBBWS9ZS5Di', arguments='{"city": "Paris"}', name='get_weather')]
ToolCallSummaryMessage -> The weather in Paris is 75 degrees.

tests/models/test_sk_chat_completion_adapter.py still passes with the patch (10 passed), and ruff check / ruff format --check are clean.

• reflect_on_tool_use=True: the follow-up create() carries toolUse/toolResult blocks but no tools, and Converse requires toolConfig whenever those blocks are present. The default (False) avoids it. Unhandled Exception while running Swarm Team with AWS Bedrock Nova models using semantic Kernel Adapter #6480 (Swarm handoffs) is in the same area.
• Bedrock's OpenAI-compatible endpoint returned {"error": {"type": "not_found_error", "code": "model_not_found", ...}} for us.amazon.nova-lite-v1:0, so OpenAIChatCompletionClient(base_url=...) is not a route for Nova today.
• AnthropicBedrockChatCompletionClient worked with anthropic 0.125.0. anthropic 1.6.0, which pip resolves today (anthropic>=0.48, no upper bound), dropped the temperature parameter, so create() raises TypeError: AsyncMessages.create() got an unexpected keyword argument 'temperature'.