Autogen Model Clients And Providers

Operator instantiates OpenAIChatCompletionClient(model='custom-llama-v1') against a self-hosted endpoint but omits model_info=. Tool calling silently fails.

For non-OpenAI-canonical model names, pass model_info={'vision': bool, 'function_calling': bool, 'json_output': bool, 'family': '...'} so AutoGen knows what capabilities the model exposes. Without model_info on unknown models, capability checks (tool calling, structured output) default conservative…

Operator wants per-token streaming in the UI. Uses model_client_stream=True on the AssistantAgent and iterates team.run_stream().

Set model_client_stream=True on the agent so model deltas surface as ModelClientStreamingChunkEvent in run_stream. UI must render chunk events as they arrive; final message arrives as a normal ChatMessage at end of stream. Don't ignore chunk events and wait for ChatMessage only — that loses the str…

Operator constructs AzureOpenAIChatCompletionClient with api_key= directly instead of Azure AD token provider.

Use azure_ad_token_provider with DefaultAzureCredential (or a workload-specific credential) so the client picks up managed-identity / workload-identity in production. Set api_version explicitly to a pinned preview/GA value. Hardcoded api_keys are fine for dev but a production smell.