LLM BASED AUTONOMOUS AGENTS FOR AUTOMATED SOFTWARE DEVELOPMENT, TESTING, DEBUGGING, AND MAINTENANCE

Abstract

The emergence of Large Language Model (LLM)-based autonomous agents represents a paradigm shift in software engineering, transitioning from passive code completion tools to active, environment-grounded cyber entities capable of orchestrating complex development workflows. This systematic review examines the architectural foundations, lifecycle applications, evaluation methodologies, and systemic trade-offs of LLM-based autonomous agents across automated software development, testing, debugging, and maintenance. We deconstruct core agent architectures comprising reasoning engines, memory structures, tool interfaces, and verification controllers that enable autonomous repository-scale operations. Our analysis reveals that multi-agent collaboration topologies introduce a fundamental trade-off between coordination dividends derived from task decomposability and context isolation and coordination taxes from handoff losses, state desynchronization, and token inflation. Empirical evaluations across benchmarks including SWE-bench Verified, SWE-bench Lite, and SWE-EVO demonstrate that while state-of-the-art models achieve resolution rates exceeding 70% on isolated bug-fixing tasks, performance dramatically declines to 18-25% on long-horizon software evolution scenarios. Notably, minimalist, scaffolded frameworks achieve competitive performance at significantly lower operational costs, challenging assumptions about required architectural complexity. We identify critical challenges including context window constraints, verification reliability, and state synchronization in extended execution trajectories. Our synthesis suggests that future advances depend on integrating neuro-symbolic verification frameworks, optimizing agent communication topologies, and developing rigorous benchmarks that better capture real-world software evolution demands. This review provides researchers and practitioners with a comprehensive framework for understanding the capabilities, limitations, and design principles of LLM-based autonomous agents in software engineering.