AutoGen

AutoGen is Microsoft’s framework for building multi-agent conversational systems. SideSeat captures runs (traces) from agent interactions.

Prerequisites

• SideSeat running locally (sideseat)
• Python 3.9+
• Model/provider credentials configured

Quick Start

1. Start SideSeat

   npx sideseat

2. Install dependencies

   pip

   pip install autogen-agentchat "sideseat[autogen]"

   uv

   uv add autogen-agentchat "sideseat[autogen]"

3. Add telemetry

   from sideseat import SideSeat, Frameworks
   from autogen import AssistantAgent, UserProxyAgent
   
   SideSeat(framework=Frameworks.AutoGen)
   
   llm_config = {"config_list": [{"model": "gpt-5-mini", "api_key": "sk-xxx"}]}
   assistant = AssistantAgent("assistant", llm_config=llm_config)
   user = UserProxyAgent("user", human_input_mode="NEVER")
   
   user.initiate_chat(assistant, message="Hello!")

4. View runs

   Open http://localhost:5388 to see your runs.

Without SideSeat SDK

1. Start SideSeat

   npx sideseat

2. Set the endpoint

   export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:5388/otel/default

3. Install dependencies

   pip

   pip install autogen-agentchat openinference-instrumentation-autogen-agentchat opentelemetry-exporter-otlp

   uv

   uv add autogen-agentchat openinference-instrumentation-autogen-agentchat opentelemetry-exporter-otlp

4. Add telemetry

   from opentelemetry import trace
   from opentelemetry.sdk.trace import TracerProvider
   from opentelemetry.sdk.trace.export import BatchSpanProcessor
   from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
   from openinference.instrumentation.autogen_agentchat import AutogenAgentChatInstrumentor
   
   provider = TracerProvider()
   provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))
   trace.set_tracer_provider(provider)
   AutogenAgentChatInstrumentor().instrument(tracer_provider=provider, skip_dep_check=True)

5. View runs

   Open http://localhost:5388 to see your runs.

Agent Configuration

Configure AutoGen agents with LLM settings:

from autogen import AssistantAgent, UserProxyAgent

config_list = [
    {"model": "gpt-5-mini", "api_key": "sk-xxx"}
]

assistant = AssistantAgent(
    name="assistant",
    llm_config={"config_list": config_list}
)

user = UserProxyAgent(
    name="user",
    human_input_mode="NEVER",
    code_execution_config={"work_dir": "coding"}
)

user.initiate_chat(assistant, message="Write a Python hello world")

Multi-Agent Conversations

Multi-agent chats are traced end-to-end:

from autogen import AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager

# Create agents
coder = AssistantAgent("coder", llm_config={"config_list": config_list})
reviewer = AssistantAgent("reviewer", llm_config={"config_list": config_list})
user = UserProxyAgent("user", human_input_mode="NEVER")

# Create group chat
groupchat = GroupChat(
    agents=[user, coder, reviewer],
    messages=[],
    max_round=10
)
manager = GroupChatManager(groupchat=groupchat)

# Start conversation
user.initiate_chat(manager, message="Write and review a sorting function")

Extracted Attributes

Attribute	Description
autogen.agent.name	Agent name
autogen.chat.round	Conversation round
autogen.message.sender	Message sender
autogen.code.execution	Code execution status

Code Execution

AutoGen’s code execution is traced:

user = UserProxyAgent(
    name="user",
    code_execution_config={
        "work_dir": "coding",
        "use_docker": False
    }
)

SideSeat shows:

• Code generation spans
• Execution spans with output
• Error spans if execution fails

What You’ll See

SideSeat shows a trace timeline with:

• A parent span for each agent conversation
• Child spans for each LLM call with model, tokens, and cost
• Code generation and execution spans with outputs
• Error spans with exception details when failures occur

Next Steps

• Python SDK — SDK reference
• Core Concepts — understanding runs, steps, and messages