Agentic Workflows in Theory and Practice

13.07.2026, 16:00 Uhr  –  Haus 9, Raum 0.17
Forschunsseminar KI und Mathematik

Johannes Zimmer (lytris)

An LLM is a conditional distribution over token sequences, not a function, so a single generation is one sample and most of what is called prompt engineering is an attempt to control a distribution by nudging its conditioning. I will make that precise, show how agentic engineering (sampling, verification, and feedback in a loop) is the practical response, and demonstrate it live: an agent builds a presentation end to end, alongside the unstructured baseline for comparison.