Colloquium Computer Science - Prof. Uwe van Heesch TH Köln - University of Applied Sciences

Title: Agentic Software Engineering and Architecture: Implications for Industry and Academia

Abstract:

AI agents are no longer confined to writing code. What began with simple code completion has evolved into agentic systems that participate across the entire software engineering lifecycle: eliciting and refining requirements, designing software architectures, making design decisions, implementing and testing, reviewing changes, and increasingly taking on planning and coordination work. This raises a fundamental question: if agents can perform much of the work traditionally done by software engineers and architects, what remains or becomes the essential human contribution?

Drawing on hands-on industrial experience and empirical research, this talk argues that software engineering and architecture do not disappear, but their center of gravity shifts from manual execution towards intent, specification, validation and oversight. Rather than implementing every aspect of a software system ourselves, we increasingly design and configure the systems that produce it. This is the emerging discipline of harness engineering: shaping the context, tools, instructions, constraints, and validation mechanisms through which agents operate. Context engineering becomes a crucial part of this work, ensuring that agents have access to the right requirements, architectural knowledge, design decisions, and project-specific information. Software architecture takes on particular significance: architectural intent, system boundaries, quality attributes, and design principles and project- or organization-specific rules must be made explicit and communicated to agents, while humans remain accountable for architectural coherence and system quality.

This shift introduces new risks. As agents take over more engineering activities, we risk accumulating cognitive debt, which is a growing gap between how our software is designed and what we genuinely understand about it. In the longer term, the delegation of engineering tasks may also lead to de-skilling, eroding the expertise needed to assess, challenge, and correct agent-generated solutions. The central challenge is therefore not to maximize agent autonomy, but to preserve intellectual sovereignty: our ability to understand, shape, evolve, and take responsibility for the systems we build.

This requires deliberately calibrating agent autonomy to find a fitting balance between efficiency gains and the human capacity to comprehend, critically assess, and remain intellectually engaged with the engineering process and its outcomes. Full autonomy cannot be the ultimate goal. These developments have far-reaching implications for both industry and academia. Established engineering processes, architecture practices, and management models must adapt to teams in which humans and agents collaborate across the lifecycle. At the same time, software engineering education faces a fundamental challenge: how do we develop and preserve engineering expertise when agents increasingly perform the activities through which that expertise has traditionally been acquired?

The talk explores these questions, presents initial approaches and solution concepts, and shares early

promising results from modernized software engineering courses at TH Köln, where students learn to work with AI agents while developing and maintaining their own engineering expertise.