AutoGen observability — drop-in. ~5 minutes.
Multi-agent conversations become navigable. Group chats, nested agents, and tool invocations land in a single trace with role attribution and message timing.
What we capture for AutoGen
- Group chat topology: speakers, turns, message order
- Per-agent system prompt and role
- Tool calls with function name, args, return value
- Conversation termination conditions and triggers
- Token + cost roll-up per agent and per conversation
- Speaker-selection decisions in the auto-selector
Install
# Python pip install trefur-observe pyautogen export TREFUR_API_KEY=trf_obs_REPLACE_ME
Instrument
from trefur_observe import TrefurObserve, patch_autogen
from autogen import AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager
obs = TrefurObserve.init(api_key="trf_obs_...")
patch_autogen(obs)
llm_config = {"model": "gpt-4o", "api_key": "<your-openai-key>"}
planner = AssistantAgent(
name="Planner",
system_message="Plan the steps to answer the question.",
llm_config=llm_config,
)
executor = AssistantAgent(
name="Executor",
system_message="Execute the plan step by step.",
llm_config=llm_config,
)
user = UserProxyAgent(
name="User",
human_input_mode="NEVER",
max_consecutive_auto_reply=4,
)
chat = GroupChat(agents=[user, planner, executor], messages=[], max_round=8)
manager = GroupChatManager(groupchat=chat, llm_config=llm_config)
user.initiate_chat(manager, message="Summarise this week's support tickets.")
# Trace shows the full speaker order and message graph in Trefur.What a AutoGen trace looks like
Three common AutoGen pain points, and what Trefur shows you
Speaker thrash
Two agents bounce a question back and forth without progress. Trefur shows the speaker timeline so you can see the loop forming and add a termination condition.
Tool calls that silently fail
An executor tool returns an empty list and the conversation continues as if everything was fine. Trefur captures the tool return so you can see what the planner actually saw.
Group-chat manager picks the wrong speaker
Auto speaker-selection sometimes picks a confused agent. Trefur captures the manager's selection logic so you can tune the system prompts or move to round-robin.
FAQ
Two-agent and group chat both?+
Both. Two-agent conversations render as a paired-span trace; group chats render as a multi-speaker timeline.
Function-calling agents?+
Yes. Registered functions and their invocations appear as tool spans inside the speaker's turn.
Code-execution agents?+
Yes. The code block and its execution result are captured as a sub-span with stdout and stderr attached.
Async AutoGen?+
Both sync and async paths emit equivalent spans. The trace structure is identical.
Ship AutoGen agents you can debug.
Free tier. No card required. First trace in under five minutes.