May 12
Loading video player...
This video introduces Agent-as-a-Graph, a novel retrieval framework designed to improve how large language models select specialized agents and tools in complex multi-agent systems. While traditional methods often struggle by searching either broad agent descriptions or isolated tool lists, this approach organizes agents and tools as interconnected nodes within a unified knowledge graph. By utilising vector search and graph traversal, the system can identify specific technical capabilities while maintaining the necessary operational context provided by the parent agent. The authors further enhance accuracy through a weighted reciprocal rank fusion (wRRF) technique that allows for fine-tuned prioritizations between agent-level and tool-level data. Rigorous testing across various embedding models demonstrates that this structured representation significantly outperforms existing state-of-the-art retrievers in both recall and ranking precision. Ultimately, the study provides a scalable solution for managing the increasing density of specialised services in autonomous AI ecosystems. #AgentAsAGraph #KnowledgeGraph #MultiAgentSystems #LLM #ToolRetrieval #AgentRouting #MCP #ModelContextProtocol #wRRF #GraphTraversal #LiveMCPBenchmark #GraphRAG