[Idea / 2024.04.14] Testing Multi-agent with supervisor architecture

[MIRACLE-cowf]

Overview

• Planning to apply a Multi-agent with supervisor architecture where one manager agents uses Tavily-search agent, Youtube-search agent, Wikipedia-search agent, arXiv-search agent, and document generation agent as team members.
• Expect time reduction, token consumption minimization, and result improvement through this.
• Please refer to the Multi-agent with supervisor link for the overall concept and flow.

Current Situation

• The code applying this architecture can be found in the AgentTeam branch. Ready For MultiAgent #3
• However, it is currently very unstable and the code is not clean.
• Initial development and prompt creation for each search engine agent completed
• Applying to the agent stage of the existing workflow.
• Currently, each agent is implemented separately in each file with LangGraph.
• Previously, LLM summarized the results of each search engine separately, but now testing results by integrating all search results and passing them to LLM.

Future Plans

• Clarify and advance the overall code and design.

• Integrate the multi-agent architecture into the existing agent workflow.

• Optimize the collaboration method between the manager agent and team member agents.

• Optimize the communication method between the manager agent and team member agents.

• Clarify and advance the roles and responsibilities of each agent.

• Compare and test the final document results with the previous version.