AI Agents Expand SWE to Six-Ring Semi-Executable Stack
AI agents introduce 'semi-executable artifacts' like prompts and workflows, expanding software engineering into a six-ring stack where outer rings—governance and societal fit—become critical engineering challenges, shifting focus from code to validation and maintenance.
Six-Ring Stack Redefines Software Engineering Scope
Researchers from Chalmers University and Volvo propose the 'Semi-Executable Stack,' a model with six concentric rings that broadens software engineering beyond traditional code. Ring 1 is executable code. Ring 2 includes prompts and natural language specs. Ring 3 covers orchestrated agent workflows. Ring 4 adds control systems like guardrails and monitoring. Ring 5 handles operational logic such as decision routines and escalation rules. Ring 6 addresses social and institutional factors, including regulations like the EU AI Act.
Historically, engineering focused on rings 1-2; now rings 2-5 demand rigorous methods, while ring 6 determines real-world viability. Execution in outer rings relies more on human or probabilistic interpretation than deterministic logic, creating 'semi-executable artifacts'—prompts, policies, workflows—that directly shape behavior but require validation. The biggest gaps are in rings 5-6, lacking mature tools compared to decades of code practices; most AI research still targets inner rings like code generation and testing.
Three observations support this: AI needs only to be 'good enough' to transform teams, not outperform top engineers; scale from everyday deployments trumps peak expertise; and domain experts building via natural language amplify the need for engineering discipline.
Developer Roles Shift to Outer-Ring Mastery
Core developer work evolves from writing code to deciding what to build, which ring to target, how to validate changes, govern them, and maintain over time. Teams using AI just for rings 1-2 gain local productivity but miss organizational redesign opportunities. Scarce skills now center on nuanced judgment in validation, governance, and upkeep, which automation makes more valuable as low-level tasks cheapen.
For instance, as domain experts create systems with natural language, engineering practices must scale to prevent chaos. This counters fears of obsolescence: AI expands the discipline, creating more engineering work in prompts, drift detection (e.g., prompt tweaks causing unexplained behavior changes), and institutional alignment.
Objections Become Solvable Engineering Tasks
Common critiques—hallucinations, reliability, messy code, maintenance—reframe as priorities. Agent hallucinations demand stronger ring 4 testing and monitoring. Faster code generation raises ring 3-5 maintenance costs. Organizational transitions turn into ring 5-6 challenges. Prompt drift exemplifies ring 2-3 issues needing versioning and traceability akin to code.
AI's impact scales through volume of small deployments, not elite performance, delivering outsized organizational value. Practitioners must engineer across the stack to capture this, treating AI as a multiplier for broader system design rather than a code accelerator.