Emir Can
I am a Senior Data Scientist/AI Engineer with a passion for building AI systems that actually work in the real world. With a background in aerospace engineering, I tend to approach AI problems as systems rather than isolated models, focusing on how different components interact, fail, and improve together.
My recent work focuses on RAG system, generative AI, and the practical challenges of making AI collaborate effectively. I am particularly interested in questions that go beyond model accuracy, such as coordination, evaluation, architecture and decision-making in complex AI systems.
Outside of my professional work, I enjoy rowing and coaching, and I actively mentor junior data and AI professionals who are looking to grow in the field. I also like sharing insights through presentations/talks and hands-on projects, with the goal of making advanced AI concepts more practical and accessible.
Session
Multi-agent LLM systems typically fall into one of two patterns: agents take fixed turns, or a single orchestrator decides who speaks next. Both are easy to implement, but both sacrifice agent autonomy. As agents become increasingly capable and specialized, productive and optimized communication between them will be crucial to unlocking their full potential. A key part of this is deciding when it is time for an agent to speak. In this talk, we will discuss whether agents can negotiate their own speaking order, and how we might evaluate their results.
We will use a Scrum refinement meeting as our multi-agent playground. A team of role-based agents, representing classic roles from a Scrum team, must collaboratively turn a business request into a backlog of implementable user stories. We implement and compare different turn-taking strategies and evaluate these using LLM-as-a-judge across metrics including participation balance, unanswered questions, redundancy, topic drift, and urgency realism.
Attendees will learn about multi-agent communication patterns, concrete turn-taking strategies they can put into practice, comparative evaluation results, a reusable evaluation strategy, and a public repository with code and notebooks they can adapt to their own multi-agent experiments.