AGDebugger is an interactive system to help you debug your agent teams. It offers interactions to:
- Send and step through agent messages
- Edit previously sent agent messages and revert to earlier points in a conversation
- Navigate agent conversations with an interactive visualization
Local Install
You can install AGDebugger locally by cloning the repo and installing the python package.
# Install & build frontend cd frontend npm install npm run build # Install & build agdebugger python package cd .. pip install .
Usage
AGDebugger is built on top of AutoGen. To use AGDebugger, you provide a python file that exposes a function that creates an AutoGen AgentChat team for debugging. You can then launch AgDebugger with this agent team.
For example, the script below creates a simple agent team with a single WebSurfer agent.
# scenario.py from autogen_agentchat.teams import MagenticOneGroupChat from autogen_agentchat.ui import Console from autogen_ext.agents.web_surfer import MultimodalWebSurfer from autogen_ext.models.openai import OpenAIChatCompletionClient async def get_agent_team(): model_client = OpenAIChatCompletionClient(model="gpt-4o") surfer = MultimodalWebSurfer( "WebSurfer", model_client=model_client, ) team = MagenticOneGroupChat([surfer], model_client=model_client) return team
We can then launch the interface with:
agdebugger scenario:get_agent_team
Once in the interface, you can send a GroupChatStart message to the start the agent conversation and begin debugging!
Citation
See our CHI 2025 paper for more details on the design and evaluation of AGDebugger.
@inproceedings{epperson25agdebugger, title={Interactive Debugging and Steering of Multi-Agent AI Systems}, author={Will Epperson and Gagan Bansal and Victor Dibia and Adam Fourney and Jack Gerrits and Erkang Zhu and Saleema Amershi}, year={2025}, publisher = {Association for Computing Machinery}, booktitle = {Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems}, series = {CHI '25} }
