Agentic Software Issue Resolution with Large Language Models: A Survey
📝 Abstract
Software issue resolution aims to address real-world issues in software repositories (e.g., bug fixing and efficiency optimization) based on natural language descriptions provided by users, representing a key aspect of software maintenance. With the rapid development of large language models (LLMs) in reasoning and generative capabilities, LLM-based approaches have made significant progress in automated software issue resolution. However, real-world software issue resolution is inherently complex and requires long-horizon reasoning, iterative exploration, and feedback-driven decision making, which demand agentic capabilities beyond conventional single-step approaches. Recently, LLM-based agentic systems have become mainstream for software issue resolution. Advancements in agentic software issue resolution not only greatly enhance software maintenance efficiency and quality but also provide a realistic environment for validating agentic systems’ reasoning, planning, and execution capabilities, bridging artificial intelligence and software engineering. This work presents a systematic survey of 126 recent studies at the forefront of LLM-based agentic software issue resolution research. It outlines the general workflow of the task and establishes a taxonomy across three dimensions: benchmarks, techniques, and empirical studies. Furthermore, it highlights how the emergence of agentic reinforcement learning has brought a paradigm shift in the design and training of agentic systems for software engineering. Finally, it summarizes key challenges and outlines promising directions for future research.
💡 Analysis
Software issue resolution aims to address real-world issues in software repositories (e.g., bug fixing and efficiency optimization) based on natural language descriptions provided by users, representing a key aspect of software maintenance. With the rapid development of large language models (LLMs) in reasoning and generative capabilities, LLM-based approaches have made significant progress in automated software issue resolution. However, real-world software issue resolution is inherently complex and requires long-horizon reasoning, iterative exploration, and feedback-driven decision making, which demand agentic capabilities beyond conventional single-step approaches. Recently, LLM-based agentic systems have become mainstream for software issue resolution. Advancements in agentic software issue resolution not only greatly enhance software maintenance efficiency and quality but also provide a realistic environment for validating agentic systems’ reasoning, planning, and execution capabilities, bridging artificial intelligence and software engineering. This work presents a systematic survey of 126 recent studies at the forefront of LLM-based agentic software issue resolution research. It outlines the general workflow of the task and establishes a taxonomy across three dimensions: benchmarks, techniques, and empirical studies. Furthermore, it highlights how the emergence of agentic reinforcement learning has brought a paradigm shift in the design and training of agentic systems for software engineering. Finally, it summarizes key challenges and outlines promising directions for future research.
📄 Content
Agentic Software Issue Resolution with Large Language Models: A Survey ZHONGHAO JIANG, The State Key Laboratory of Blockchain and Data Security, Zhejiang University, China DAVID LO, School of Computing and Information Systems, Singapore Management University, Singapore ZHONGXIN LIU∗, The State Key Laboratory of Blockchain and Data Security, Zhejiang University, China Software issue resolution task aims to address real-world issues in the software repositories (e.g., bug fixing, efficiency optimization) based on natural language descriptions provided by users, representing a key aspect of software maintenance. With the rapid development of large language models (LLMs) in reasoning and generative capabilities, LLM-based approaches have made significant progress in automated software issue resolution. On the other hand, real-world software issue resolution is inherently complex and requires long- horizon reasoning, iterative exploration, and feedback-driven decision making that demand agentic capabilities beyond conventional single-step approaches. Recently, LLM-based agentic systems have become mainstream for software issue resolution. Advancements in agentic software issue resolution can not only greatly enhance software maintenance efficiency and quality but also provide a realistic environment for validating agentic systems’ reasoning, planning, and execution capabilities, bridging AI and software engineering. This work presents a systematic survey of 126 recent studies at the forefront of LLM-based agentic software issue resolution research. It outlines the general workflow of the task and establishes a taxonomy across three dimensions: benchmarks, techniques, and empirical studies. Furthermore, it highlights how the emergence of agentic reinforcement learning has brought a paradigm shift in the design and training of agentic systems for software engineering. Finally, it summarizes key challenges and outlines promising directions for future research. The artifacts’ page accompanying this survey is at https://github.com/ZhonghaoJiang/Awesome- Issue-Solving. 1 Introduction In software engineering, software maintenance often accounts for about two-thirds of software lifecycle costs [95, 183]. Software issue resolution occupies a crucial position in software mainte- nance [87, 93]. This task aims to understand, locate, and resolve issues in real-world code repositories based on developers’ natural language descriptions of issues, covering diverse maintenance activi- ties such as bug fixes, feature additions, and efficiency optimizations. Hereafter, we refer to this task as issue resolution for brevity. Traditional issue resolution methods heavily rely on human expertise, making the process time-consuming, error-prone, and difficult to scale [128]. Consequently, issue resolution has long been regarded as a critical bottleneck to efficient software evolution [19]. In recent years, LLMs have achieved significant success in multiple areas of software engineer- ing [10, 20, 165] due to their rapid advancements in code understanding [138], reasoning [57, 162], and generation [73], paving the way for automated issue resolution. However, issue resolution in real-world software systems is inherently complex, typically involving long-horizon reasoning, iter- ative exploration, interaction with evolving codebases, and feedback-driven decision making, which goes beyond the capabilities of conventional single-step automated approaches. LLM-based agentic systems are autonomous, goal-driven AI architectures that can interpret objectives, plan multi-step tasks, and adapt their behavior based on environmental feedback [18]. They have become the mainstream and state-of-the-art approach for issue resolution [157, 164, 174]. LLM-based agentic ∗Corresponding author. Authors’ Contact Information: Zhonghao Jiang, zhonghao.j@zju.edu.cn, The State Key Laboratory of Blockchain and Data Security, Zhejiang University, Hangzhou, China; David Lo, davidlo@smu.edu.sg, School of Computing and Information Systems, Singapore Management University, Singapore; Zhongxin Liu, liu_zx@zju.edu.cn, The State Key Laboratory of Blockchain and Data Security, Zhejiang University, Hangzhou, China. arXiv:2512.22256v1 [cs.SE] 24 Dec 2025 2 Jiang et al. issue resolution tools have been widely adopted in real-world software development. For example, as of May 2025, Trae [51], one of the most powerful agentic issue resolution and code assistance tools, has exceeded 1,000,000 monthly active users and has helped deliver over 6 billion lines of code [8]. On the other hand, because of the complexity and broad applicability, issue resolution has become a key task for evaluating LLMs and software engineering (SE) agentic systems [11, 68, 184]. Advancing LLM-based agentic issue resolution not only holds the potential to significantly improve the efficiency and quality of software maintenance but also provides LLM-based agentic systems with a realistic environment to
This content is AI-processed based on ArXiv data.