Coding Agents: A Comprehensive Survey of Automated Bug Fixing Systems and Benchmarks

One of the trickiest problems in software engineering is automating software issue fixes, which calls for a thorough comprehension of contextual relationships, code semantics, and dynamic debugging techniques. The development of automatic program repair (APR) is examined in this survey, which traces a path from early template and constraint-based approaches to more recent developments powered by large language models (LLMs). Three main paradigms are compared here: retrieval-augmented approaches that integrate external knowledge sources, agent-based systems that use multi-agent frameworks, and agentless systems that use simplified repair pipelines. Real-world benchmarks that mimic actual engineering workflows and repository-level difficulties, such as SWE-bench, CODEAGENT-BENCH, and CodeRAG-Bench, are used to assess these cutting-edge technologies. This study demonstrates how agentic, agentless, and retrieval-augmented systems use LLMs to achieve previously unheard-of precision and scalability by following the shift from localized, single-file solutions to solving complicated, multi-file, and repository-wide difficulties. According to our findings, while complex agent architectures have potential, straightforward test-time scaling frequently produces better outcomes, especially when paired with containerized environments that allow for parallel exploration. Additionally, the survey looks at industrial applications, emphasizing effective connections with quality assurance and DevOps procedures. In order to further the development of more resilient and flexible APR frameworks that blend in perfectly with contemporary software engineering practices, we conclude by highlighting important issues in context handling and validation and suggesting future research directions in improved contextual models, human-AI collaboration, and multi-modal debugging systems.