Scaling Validation: The Backbone of Agentic Software Engineering

Wann: 03. November 2026, 16:00 Uhr; anschließend Zeit zur Diskussion

Wo: paluno, SGW-009, Gerlingstraße 16, 45127 Essen

Anmeldung: Die Veranstaltung ist kostenfrei. Da die Raumkapazität begrenzt ist, empfehlen wir eine frühzeitige Anmeldung unter:

https://terminplaner6.dfn.de/b/987d4a4c69be048af9d16e45366e4136-1929097 
 

Abstract:

Coding agents now plan changes, edit whole repositories, run tests and open pull requests. As a result, the scarce resource in software engineering is shifting: producing code has become cheap, while establishing confidence in that code has become more expensive. Industry surveys name review and validation of AI-generated code as the main bottleneck in Agentic Software Engineering. Moreover, studies show that LLM-based approaches cannot solve this validation challenge on their own.

This talk argues that the central research question of today's Agentic Software Engineering is how to efficiently improve the quality of generated code through symbolic feedback to coding agents. What an agent can be trusted to deliver is bounded by the strength of the validation signals in its loop. Tests are cheap but narrow, and agents have been shown to occasionally circumvent them. Formal proof is strong but expensive, and demands correct specifications. Static analysis sits in between: It is semantic, property-driven and independent of the agent. Yet, while static analysis is actively used as a guardrail in coding agents' inner loops, false positives in the analysis cause agents to chase spurious warnings (burning tokens), and false negatives give false assurance. Slow-running analyses can greatly delay agentic loops.

I will present SOSA (Self-Optimizing Static Analysis), an ERC-funded effort to build analyses that configure and adapt themselves to the agent's task, the codebase and the time budget at hand. The goal is to deliver the strongest signal possible for the given situation, thus increasing the efficiency of Agentic Software Engineering.

Bio:

Eric Bodden is one of the leading experts on secure software engineering, with a specialty in building highly precise tools for automated program analysis. He is Professor for Secure Software Engineering at Paderborn University and director for Software Engineering and IT-Security at Fraunhofer IEM, where he is collaborating with the leading national and international software development companies. Bodden is a member of Acatech, der German National Academy of Science and Engineering, as well as a member of the review board on Software Engineering and Programming Languages of the German Science Foundation (DFG).

Prof. Bodden's research was awarded numerous times. For instance, in 2024 he received an ERC Advanced Grant, and in 2019, Bodden was named ACM Distinguished Member. At the German IT-Security Price, his group scored 1st place in 2016 and 2nd place in 2014. In 2014, the DFG awarded Bodden the Heinz Maier-Leibnitz-Preis, Germany's highest honour for young scientists. In 2020, Bodden was elected into the Working Group 2.4 Software Implementation Technology of the International Federation for Information Processing (IFIP).

Personal website:  https://www.uni-paderborn.de/person/59256