The Semi-Executable Stack: Agentic Software Engineering and the Expanding Scope of SE
New framework redefines software engineering for the AI era, moving beyond just code to probabilistic systems.
A new academic paper by Robert Feldt and colleagues reframes the perceived threat of AI to software engineering as a fundamental expansion of the discipline's scope. The authors argue that AI agents, powered by LLMs like GPT-4 and Claude 3 and capable of multi-step planning, are shifting the engineering target from purely deterministic, executable code to what they term the 'Semi-Executable Stack.' This stack encompasses a broader set of artifacts, including natural language instructions, tool-using workflows, control mechanisms, and organizational routines, whose execution relies on probabilistic interpretation rather than strict determinism.
To navigate this shift, the paper introduces a six-ring diagnostic reference model. The rings span from core executable artifacts out to societal and institutional fit, providing a framework for teams to identify where contributions, bottlenecks, or necessary organizational transitions lie. The authors develop their argument through three worked cases and reframe common objections as engineering challenges. They conclude with a 'preserve-versus-purify' heuristic, a practical guide for deciding which traditional software engineering processes and controls to maintain versus which to simplify or redesign for the new AI-augmented reality.
- Introduces the 'Semi-Executable Stack,' a six-ring model for engineering AI-augmented systems that include probabilistic elements.
- Argues AI agents (LLM-driven systems) expand software engineering's scope to workflows and controls, not just executable code.
- Provides a 'preserve-versus-purify' heuristic for organizations to adapt legacy engineering processes for the AI era.
Why It Matters
Provides a crucial framework for engineering teams to systematically integrate AI agents and manage the transition to probabilistic, AI-driven systems.