AGENTIC ARTIFICIAL INTELLIGENCE FOR AUTONOMOUS SOFTWARE ENGINEERING: INTEGRATING CODE REASONING, FORMAL VERIFICATION, AUTOMATED TESTING, AND REPOSITORY-LEVEL PROGRAM EVOLUTION

Abstract

The software engineering discipline is undergoing a structural paradigm shift from passive generative code autocompletion toward autonomous, agentic artificial intelligence (AI) systems capable of repository-scale software development. Early Large Language Model (LLM) implementations operated as stateless text-to-code engines, suffering from fundamental failure modes including inability to navigate non-local dependencies, absence of execution feedback awareness, and susceptibility to logically flawed code generation. Agentic AI addresses these bottlenecks by embedding reasoning engines within interactive control loops, enabling autonomous agents to perceive environment states, decompose abstract goals into structured plans, invoke external tools, and iteratively refine actions based on dynamic feedback. This review synthesizes the theoretical foundations, system architectures, verification mechanisms, testing paradigms, and socio-economic impacts of agentic AI in software engineering across four interconnected pillars: agentic code reasoning and control architectures formalized as Partially Observable Markov Decision Processes; mechanized formal verification coupling neural generation with Interactive Theorem Provers and SMT solvers; automated testing and execution feedback loops deploying coverage-guided fuzzing and mutation-based test synthesis; and repository-level program evolution scaling repairs from single-issue fixes to multi-commit release evolution. We examine comparative scaffold paradigms spanning model-driven, procedure-based, and self-evolving architectures. We analyze critical vulnerability surfaces including indirect prompt injection, the self-repair trap, and cognitive debt. We further evaluate enterprise governance frameworks, economic make-or-buy bifurcation, and benchmark evolution from HumanEval through SWE-bench to SWE-EVO. Finally, we outline future research horizons toward provably correct, autonomous software development, emphasizing that ultimate accountability remains strictly human.