Andreas Hermann

I work on evaluating AI systems as they're actually deployed — agents acting through tools, over time, with people supervising them — where the failures don't show up in single-model evals. That includes how agent behaviour changes in composition, and what makes an agent's own account of what it did trustworthy or not. My background is a PhD in Information Systems, studying how people act inside socio-technical systems: 17 peer-reviewed papers, 450+ citations, and a €300K EU project I led. Alongside that, years of keeping machine-learning systems running in front of real traffic. I co-authored an empirical study of alignment differences between single-agent and multi-agent LLM systems and what they imply for human–AI teaming. From August 2026 I'm on AI safety full-time.

Personal Website

https://www.ahermann.org/

LinkedIn

LinkedIn

Projects

No grants recorded.