This paper is the write-up of a keynote titled “Agentic Software Engineering Will Eat the World: AI-Based Systems as the New Operating System of Society”, given at the Agentic Engineering 2026 workshop in Rio de Janeiro on April 14, 2026; its authors are affiliated with Chalmers University of Technology and Volvo Group. It starts from the anxiety that stronger foundation models and tool-using agents make tasks such as scaffolding, routine test generation, straightforward bug fixing and small integration work look exposed, and that experienced practitioners worry their expertise will lose value. The authors take those concerns seriously but draw a different conclusion: agentic systems do not make software engineering smaller, they expand the engineering object outward, from code and tests toward prompts, workflows, controls, organizational operating logic and eventually societal and institutional fit.
The paper’s central concept is the Semi-Executable Artifact — an artifact that helps specify, coordinate or constrain a software-intensive system’s behavior but whose enactment depends on interpretation by humans, probabilistic models or both. Around it the authors build the Semi-Executable Stack, a six-ring reference model whose purpose is diagnostic: to show which ring a contribution or bottleneck primarily inhabits and which adjacent rings it depends on. The paper describes itself as diagnostic and agenda-setting rather than empirical, and explicitly not a maturity model, a complete process framework or an evaluation rubric.
Key Points
- The paper argues that three observations push the engineered object outward even before any agentic system matches expert engineers: organizational change is triggered by adequacy against how scarce expert attention is actually allocated, not by parity with the best human engineer; thousands of low-friction uses compound into capability that sporadic access to a top expert cannot match; and lowering the barrier to creation, as end-user software engineering showed, raises rather than lowers the demand for engineering discipline.
- The stack’s six rings are executable artifacts (code, tests, schemas, configurations); instructional artifacts (prompts, natural-language specifications, task descriptions, exemplars); orchestrated execution (tool use, retrieval, agent workflows, multi-agent protocols, human-agent loops); control systems (guardrails, monitoring, evaluation harnesses, policy layers, escalation rules); operating logic (decision preparation, coordination routines, delegation structures); and societal and institutional fit (regulatory fit, institutional legitimacy, cross-organizational integration).
- The rings mark regions on a spectrum of specification completeness and execution determinism, with blurry boundaries; the authors state that they are not lifecycle phases, maturity levels or runtime layers. Historically software engineering centered on ring 1 and partly ring 2; the claim is that rings 2–5 increasingly become first-class engineering objects while ring 6 increasingly constrains what succeeds.
- A contribution’s primary ring is identified by asking whether the system’s value would collapse if that ring’s artifact were removed or poorly designed. The paper pairs the rings with a behavioral question — whether a change mainly affects individual cognition, team coordination or organizational routines — treating behavioral software engineering as a cross-cutting lens rather than a seventh ring.
- Three worked cases illustrate the diagnosis: an automated automotive API-testing pipeline sits mainly at rings 2–3, an LLM-based multi-agent system for automotive release go/no-go decisions sits at ring 5 supported by rings 3–4, and a field experiment on AI and teamwork outside software sits at ring 5. As the primary object moves outward, the authors argue, the evaluation criteria, required skills and unresolved bottlenecks shift with it.
- Five common objections are reframed as engineering moves: reliability failures (rings 1–3) make verification and oversight core artifacts; maintenance debt (rings 2–4) calls for versioning prompts and workflow logic and tracking evaluation suites alongside code; organizational inertia (rings 4–5) makes adoption an engineering concern; power and politics (rings 4–6) make governance one; and the judgment that resists automation (rings 5–6) moves the center of gravity outward. The authors add that where failures would be catastrophic and irreversible, or the needed controls cannot yet be built, some deployments should not happen yet.
- The preserve-versus-purify heuristic keeps durable principles — explicit reasoning about requirements, modularity and interfaces, validation and verification, traceability, lifecycle thinking and socio-technical realism — and purifies practices that mainly compensated for older constraints, such as those optimized for manual code production, process artifacts built for low-bandwidth human coordination, and strict phase boundaries inside interactive human-agent loops.
- For researchers, the paper names SE4AI work at rings 5 and 6 — decision routines, operating logic, governance and institutional fit — as the most underserved frontier. For practitioners it suggests a diagnostic: name the ring whose failure would kill a project and the adjacent rings it depends on, and compare them with the ring receiving most of the engineering effort; when they diverge, the authors say, projects tend to stall.
Notes
The authors position the stack alongside several literatures rather than replacing them: AI4SE surveys and agentic-coding benchmarks such as SWE-Bench, which they place mainly in rings 1–3; MLOps and holistic evaluation work at ring 4; socio-technical and behavioral software engineering; and work on AI governance, including the EU AI Act and NIST’s AI risk-management framework, at rings 5 and 6. They also point to developer systems such as Claude Code, where prompts, repository instructions, tool permissions and reusable skills partly encode how work should proceed, as evidence that the shift is already visible in practice.
Limitations the authors state: the stack is a reference model rather than an empirical result; two of the three cases share an automotive industrial context and the third demonstrates portability rather than full generalization; the evidence base for the outer rings is thinner, so claims there are directional; and ring assignment is often ambiguous, with the stack meant to support that discussion rather than adjudicate it.