Workshop Program

13:30 - 13:40	Workshop Opening
13:40 - 14:10	Keynote From Hype to Engineering Reality — Navigating GenAI Adoption in Software Projects Zoran Roncevic, Adnovum Generative AI is quickly becoming part of everyday software engineering practice, yet many organizations struggle to turn early experimentation into sustainable engineering capability. This keynote examines the gap between GenAI hype and production reality, focusing on architectural, governance, and quality challenges observed in real-world projects. Drawing on hands-on experience from large-scale software environments, the talk shows how architects can guide responsible adoption through clear boundaries, integration patterns, and validation mechanisms. It offers practical guidance on balancing productivity gains with maintainability, reliability, and long-term system integrity when GenAI becomes part of the engineering toolchain, across teams and evolving delivery constraints.
14:10 - 14:35	Presentation of accepted contribution LLM-Based Agentic Systems for Software Engineering: Challenges and Opportunities Yongjian Tang and Thomas Runkler (Siemens AG, Technical University of Munich)
14:35 - 15:00	Presentation of accepted contribution Can LLMs Fool Themselves? Roman Machacek and Timo Kehrer (University of Bern)
15:00 - 15:30	Coffee Break

15:30 - 16:05	Keynote Designing Agentic Loops that Actually Ship Chirag Agrawal, Amazon Web Services Agentic AI systems are closed loop programs that plan, invoke tools, evaluate results, and iterate until a goal is met, yet there is great deal of depth involved in running them sustainably in production. This session presents a practical framework for designing that loop for production. We focus on the ReAct as the foundation for tool-based agents and multi-agent collaboration, disciplined tool selection and reliable routing, and the operational realities of deploying agents at scale. We address context rot in multi-turn conversations and share robust context management practices that keep agents on track. We also cover cost control in multi-turn systems and techniques that improve perceived latency, including prompt caching and context compression. Attendees leave with a concise checklist, actionable patterns, and guardrails they can apply immediately to ship resilient agentic features.
16:05 - 16:30	Presentation of accepted contribution Towards neuro-symbolic constrained decoding for reliable code generation with LLMs Philipp Kogler, Wei Chen and Alois Haselböck (Siemens Österreich)
16:30 - 17:00	Reflexion round & end of workshop