SERG Seminar

In this edition of our weekly SERG seminar, we will hear from:

Roham Koohestani on “When Agents Stop Being Programs: Rethinking Assurance Paradigms for Agentic AI”

Abstract:
Modern agentic AI systems act autonomously over long horizons in stochastic environments. These properties render traditional deterministic verification techniques insufficient. This talk presents a formal perspective on agentic AI grounded in trace semantics and probabilistic automata. We begin by modeling agents as stochastic transition systems with explicit memory disciplines, which yields clear decidability boundaries for verification. Building on this foundation, we introduce Dynamic Probabilistic Assurance, a runtime verification paradigm that learns behavioral models online and computes quantitative guarantees. We then present Tri-CEGAR, a trace-driven abstraction and refinement mechanism that enables scalable probabilistic model checking without manual state engineering. Finally, we discuss how these ideas can be operationalized in real-world agent frameworks through tools such as AgentGuard, and outline future directions involving communicating probabilistic automata and multi-agent assurance.

If you are interested to give a talk or host a discussion session in one of our next meetings, please contact Carolin Brandt via Mattermost or email.