Agentic workflows for end-to-end software engineering automation

Abstract

With the rise of large language models (LLMs) and independent AI systems, software engineering is undergoing a transformational shift. This conceptual paper theorizes agentic workflows systems, where an AI agent or agents proactively perceive, plan, act and reflect throughout the entire software development lifecycle (SDLC); its implications for end to end software engineering automation are also discussed. This paper constructs a theory for agentic SE systems viewed from four different angles: architectural configuration, reasoning capability, SDLC coverage, and evaluation validity, referring to 30 basic and contemporary papers from the past 25 years since 2000. It is then supported with quantitative evidence from benchmark studies: top agentic systems can now solve up to 43% of real-world GitHub issues on SWE-bench Lite, with 85.9% Pass@1 on Human-Eval and 55.8% less time spent on completing a developer task in controlled experiments. There are, however, significant theoretical tensions that are not resolved: autonomy and oversight, benchmark performance and validity in the real world, and the capability of the system and ethical responsibility. Finally, the paper outlines a research agenda that focuses on the specification level of agentic SE systems, on their self-verification, and on their governance.