Panel Discussion
Panel Discussion
Data Harness for Agentic LLMs
Recent advances in the agentic LLM system have largely treated the data layer as a passive service that returns top-k documents or executes a single-shot text-to-sql. In practice, this reduces databases and search engines into a passive data provider, analogous to pressing the "I'm Feeling Lucky" button in Google for every query. However, researchers and practitioners trained in database systems know that a data harness is far richer than the raw data it stores: schemas encode semantic structure, query optimizers expose exploration over large plan spaces, and execution engines produce valuable intermediate signals such as selectivity, cardinalities, and join paths. This panel will brainstorm what should be the role of a data harness in new problem and solution spaces.
Moderator
Panelists
-
Victor Leis (TUM) -
Wenjie Zhang (University of New South Wales) -
Meihui Zhang (Beijing Institute Technology)