Collaborative Human–Agent Interaction: Balancing Human Interruptions and Feedback

This is the codebase for the my Bachelor's thesis at the IVIA Lab, ETH Zurich. It comprises a visual interface that allows a user to interrupt an AutoGen agent team in real-time, provide feedback for it (human-initiated interruption) and at the same time be prompted by the agents for feedback (agent-initiated feedback). This involved modifying the AutoGen library, implementing a specific interface design and connecting the agents to tool functions, including MCP servers.

Backend

• backend/: Contains all the logic to build an AutoGen agent team dynamically, and stream it to the interface frontend via a WebSocket + FastAPI setup

See Sections 5.3, 5.4 & 5.5

Frontend

• frontend/: Visual interface for multi-agent communication of AutoGen agent teams. Compatible only with the 2 specific agent teams used in the thesis

See Chapter 6

MCP

• mcp/: Directory that harbors all the MCP servers that the AutoGen agents have access to. In our case, it is the Congress MCP server under the name ragMCP

See Section 5.2

How to Run

1. Make sure that you have all the .env/secrets files filled up with your API keys. The following files are crucial

• .env
• backend/.env
• mcp/secrets.ini

Detailed instructions on how to obtain the API keys are found in the subdirectories

2. (Optional) Choose the team setup you wish to run in backend/factory/team.yaml. The Congress team is in backend/factory/team.yaml.congress_backup and the Deep Research team is in backend/factory/team.yaml.research_backup.

3. Start the Docker container:

docker compose up --build

4. Navigate to localhost 5173 and play around with the site!

Thesis

In order to read the thesis, click here.