AI Agents Expand Software Engineering Beyond Code, Researchers Say
Researchers at Chalmers University of Technology and the Volvo Group challenge the idea that AI agents will make software developers obsolete. In a recent paper, they state these systems actually widen the scope of software engineering. They introduce the concept of "semi-executable artifacts," which cover items such as prompts, workflows, policies, escalation rules, and decision routines. These elements influence system actions much like traditional code does, though they depend on human judgment or chance for execution.
Chalmers University of Technology, based in Gothenburg, Sweden, has long focused on engineering and computer science research since its founding in 1829. The Volvo Group, a major Swedish automotive company headquartered nearby, often collaborates with the university on industrial projects.
The Semi-Executable Stack Model
The paper presents a key framework named the "Semi-Executable Stack," structured as six concentric rings. Ring 1 at the core contains traditional executable code. Ring 2 includes prompts and specifications written in natural language. Ring 3 involves workflows that orchestrate multiple AI agents.
Ring 4 consists of control mechanisms, such as guardrails and monitoring tools. Ring 5 covers organizational logic for operations, including decision-making routines. The outermost Ring 6 addresses social and institutional factors, like the EU AI Act.
This stack shows how software engineering now spans beyond just code. Execution in outer rings relies more on human interpretation than on strict machine processes. Historically, the field emphasized rings 1 and 2. Today, rings 2 through 5 demand more engineering attention, while ring 6 often determines real-world success.
Gaps and Priorities in Current Practices
The researchers identify the largest shortcomings in rings 5 and 6. Decades of methods exist for code in inner rings, including generation, bug fixes, testing, and benchmarks. Yet, techniques for decision routines, governance, and institutional alignment remain underdeveloped. Most studies still target rings 1 through 3.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
They support their view with three key points. First, AI does not need to outperform top engineers to alter team dynamics; being adequate suffices. Second, widespread use of modest AI tools provides greater organizational value than occasional elite expertise. Third, as non-experts create systems via natural language, demand for solid engineering rises.
Reframing Criticisms as Engineering Challenges
Common concerns about AI agents, such as unreliability or poor code quality, get recast as solvable problems. Hallucinations by agents call for stronger testing and monitoring. Faster code production raises maintenance demands proportionally.
Issues like "prompt drift," where small changes to prompts alter behavior unpredictably, highlight new needs. Organizational adaptations during this shift become engineering tasks themselves. Nuanced decisions resist automation, making them more critical as routine work automates.
Evolving Skills for Developers
For working professionals, the paper stresses a clear change: rare expertise now lies in choices about what to construct or modify. This includes selecting the relevant ring, validation methods, governance approaches, and long-term upkeep.
Teams viewing AI solely as a boost for rings 1 and 2 might gain short-term speed. However, they overlook broader organizational changes.
The work ties to a keynote by Robert Feldt at the Agentic Engineering 2026 Workshop in Rio de Janeiro. It also builds on automotive industry efforts with Volvo partners.