What does it take for AI systems to truly understand code and what happens when software is no longer designed only for humans, but for AI agents too?

Join Munich🥨NLP and JetBrains for an evening exploring Code Intelligence and AI Agents, bringing together research and industry perspectives on how LLMs understand program behavior and how humans and coding agents can collaborate on software development.

Talk 1: Code World ModelsTobias Lindenbauer · JetBrains Research

Talk 2: Designing Software for Humans and AgentsArtem Kislovskiy · Artefact

Two talks, pizza, drinks & snacks, and plenty of time to connect with Munich’s NLP, AI, and developer community.

📅 Agenda

17:45 | Doors open

18:20 | Pizza arrives 🍕

18:50 | Welcome & IntroJetBrains × Munich🥨NLP

19:00 | Talk 1: “Code World Models”Speaker: Tobias Lindenbauer · JetBrains Research

Abstract

Modern LLMs perform well at generating code that solves a given task, but they may introduce regressions elsewhere in the codebase, or simply write code with poor runtime or memory performance.

This has led researchers to investigate how well LLMs model program behavior internally: Can they tell whether a test passes or fails without running it? Do they have a feel for which parts of the code to target to robustly optimize runtime or memory?

In this talk, Tobias will outline the emerging Code World Models research direction and share early findings from an ongoing project at JetBrains Research.

About the Speaker

Tobias Lindenbauer is a Research Engineer at JetBrains Research working on Code World Models for improved LLM code understanding.

Previously, he worked on context management, LLM routing, memory systems, and benchmarks, resulting in publications at DL4C@NeurIPS and REALM@ACL.

19:40–19:50 | Short break

19:50 | Talk 2: “Designing Software for Humans and Agents”Speaker: Artem Kislovskiy · Artefact

Abstract

One coding agent writes; another challenges.

In this live demo, Claude Code and Codex share a split terminal in an actor–critic loop: the actor implements changes, while the critic reviews them, questions assumptions, and runs checks. Feedback drives the actor’s next revision, and the critic checks again.

We’ll explore how clear responsibilities, inspectable state, and verifiable actions support this collaboration—with the human deciding who acts and who critiques.

About the Speaker

Artem Kislovskiy is a Senior Software Engineer at Artefact in Lausanne, building reliable, Swiss-made software.

20:30 | Networking, drinks & snacks

🎟️ Registration

Registration is free but subject to host approval due to limited capacity. If your plans change, please cancel your registration so we can offer the spot to someone else.