AUTONOMOUS SOFTWARE ENGINEERING WITH LARGE LANGUAGE MODELS SELF-HEALING CODE GENERATION, VERIFICATION, AND CONTINUOUS PROGRAM REPAIR

Abstract

The rapid evolution of Large Language Models (LLMs) has catalyzed a fundamental paradigm shift in software engineering, transitioning from human-centric development workflows to autonomous agentic systems capable of self-directed code synthesis, verification, and continuous repair. This comprehensive review examines the foundational mechanisms enabling autonomous software engineering, synthesizing recent advances across four interconnected domains: cognitive multi-agent architectures, self-healing code generation mechanics, execution-grounded verification frameworks, and continuous program repair within CI/CD pipelines. We analyze the evolutionary trajectory from tool-augmented assistants to self-evolving multi-agent ecosystems, highlighting how contemporary frameworks such as AgentForge and STORM leverage role-specialized coordination and workspace state management to overcome context window limitations and error accumulation. The review critically evaluates iterative self-repair dynamics, demonstrating that modern instruction-tuned models achieve significant performance improvements through error-guided feedback loops, with repair efficacy varying substantially across error types—name errors (77%), syntax errors (66%), and logical assertion failures (45%). We examine constraint-based program synthesis methodologies including PathFix, which employs path-sensitive constraint extraction and SMT-solving to eliminate patch overfitting. The security implications of autonomous patch generation are systematically analyzed, revealing that LLM-generated patches introduce vulnerabilities at rates orders of magnitude higher than human developers, necessitating multi-layered governance frameworks incorporating static analysis, dynamic probability filtering, and mandatory human-in-the-loop approval. Finally, we survey the evolving benchmark landscape from SWE-bench Verified to SWE-bench Pro, demonstrating significant performance degradation on proprietary, contamination-resistant codebases. This review provides essential insights for researchers and practitioners navigating the transition toward trustworthy autonomous software engineering systems.