Agentic AI
editOrganizers:
Room: Turchese
09:30-09:40 Welcome and introduction
09:40-10:00 Agents for Scientific Discovery: A Parallel Universe?
10:05-10:25 Building Generalist Humanoid Robots
10:30-10:50 What is a useful action?
11:00-11:30 Coffee Break
11:30-11:50 Open Source LLMs for the Research Community
11:50-13:00 Hands-on Agentic AI Workshop and Group Discussion
13:00-14:30 Lunch Break
14:30-14:50 Talk by Brent Mittelstadt
14:50-15:10 DOAgent: A Data-Oriented Library for Observable Multi-Agent Systems
15:15-16:00 Panel Discussion
16:00-16:05 Closing Remarks
Abstracts
Agents for Scientific Discovery: A Parallel Universe?
Boris BollietWe will present how multi-agent systems can be designed to perform scientific research. We will show how we are using them to solve hard open-ended research problems, in some cases outperforming humans operating without them. The agents and systems we develop act as a force multiplier and generate ideas that are unknown to us, but it is important to emphasize that they do not operate in isolation. They augment the humans; they do not replace them. We will then lay out a vision of self-organising AI scientists operating in a parallel cyber space, co-evolving science alongside humans.
Building Generalist Humanoid Robots
Yuke ZhuIn an era of rapid AI progress, leveraging accelerated computing and big data has unlocked new possibilities to develop generalist AI models. As AI systems like ChatGPT showcase remarkable performance in the digital realm, we are compelled to ask: Can we achieve similar breakthroughs in the physical world — to create generalist humanoid robots capable of performing everyday tasks? In this talk, I will present our data-centric research principles and approaches for building general-purpose robot autonomy in the open world. I will discuss our recent works leveraging real-world, synthetic, and web data for training robotic foundation models. By combining these advances with cutting-edge developments in humanoid robotics, I will outline a roadmap for the next generation of autonomous robots.
What is a useful action?
Özgür ŞimşekHow can autonomous agents develop useful action hierarchies on their own? This research question is critically important for achieving fully autonomous behaviour in large, complex environments. In this talk, we will explore two promising approaches. We will first explore the graphical representation of the agent-environment interaction, focusing on how modularity maximisation can expose the temporal structure of this interaction graph at multiple levels of abstraction, enabling the creation of action hierarchies that closely match human intuition and improve learning performance. We will then explore, as an alternative complementary approach, agent behaviour that led to successful outcomes in the past, focusing on how recurring action sequences can be brought together to form new actions that enable rapid adaptation to new tasks.
Open Source LLMs for the Research Community
Ryan DanielsModern AI research faces a crisis of dependency. As we increasingly rely on commercial APIs and closed-source models, the scientific community risks undermining the reproducibility, interpretability, and sustainability of its work. This talk argues for a pivot: moving from viewing AI as a tool we rent, to viewing it as infrastructure we control. We present a case study in building this "sovereign stack." We demonstrate how we architected a server to operate as a shared utility for the academic community. But independence brings complexity. We candidly explore the significant engineering friction encountered in moving from raw hardware to a production-ready service. We discuss the realities of hardening the attack surface, managing API keys, and the intricate balancing act of optimizing throughput, latency, and VRAM usage against massive context windows. By releasing our full software stack, monitoring configurations, and Architecture Decision Records (ADRs), we aim to "outsource the tedium" of these discoveries. This talk serves as a blueprint for how institutions can build secure, scalable, and open AI infrastructure that keeps science reproducible and data secure.
Talk by Brent Mittelstadt
Brent MittelstadtTBD
DOAgent: A Data-Oriented Library for Observable Multi-Agent Systems
Christian CabreraIntellectual debt emerges when we deploy ML-based systems we do not fully understand and control. Multi-agent systems (MAS) can generate intellectual debt when agents make opaque decisions. DOAgent is a data-oriented library that addresses this by making agent behaviour observable and traceable through a shared data model. This talk introduces the library, demonstrates how it supports diverse policies from heuristics to LLMs, and shows how we can build MAS that are observable and interpretable by design.