Why Do Multi-Agent LLM Systems Fail?
Join the Bilkent AI Society’s 11th Seminar Featuring Mert Cemri (UC Berkeley)
Title: Why Do Multi-Agent LLM Systems Fail? A Deep Dive into MAST and the Future of AI Agents
We are thrilled to host Mert Cemri, PhD researcher at UC Berkeley and co-author of the influential paper “Why Do Multi-Agent LLM Systems Fail?”, for a thought-provoking session on the hidden challenges of building large language model (LLM)-based multi-agent systems.
Despite the growing hype around agent-based AI, recent evidence shows that these systems often fail more than they succeed. Why? In this talk, Mert will unveil MAST (Multi-Agent System Failure Taxonomy), the first empirically grounded framework to systematically understand why multi-agent systems underperform.
Learn how over 200 real-world MAS traces were analyzed to extract 14 failure types, categorized under specification issues, inter-agent misalignment, and verification breakdowns. Discover how even top-tier systems like ChatDev and MetaGPT struggle with subtle coordination bugs, overlooked specifications, and flawed verification processes.
The talk will also feature insights from the development of an LLM-as-a-judge evaluation pipeline, and how MAS failures relate more to organizational design than model capabilities, a critical shift in thinking.
🔗 Who Should Attend:
AI and LLM researchers interested in multi-agent architectures
Engineers and builders working on AI agents and orchestration
Researchers focused on evaluation, trustworthiness, and AI failure analysis
Anyone exploring the future of robust and scalable AI systems
📅 Date: June 19, 2025
🕔 Time: 19:00 TRT (GMT+3)
📍 Location: Online (Register to get the meeting link)
🌐 Register now for free: https://register.bilkentai.com
This session promises to redefine how we think about AI agents, from architecture to accountability. Don’t miss this chance to engage with cutting-edge research shaping the future of intelligent systems.