Autonomous Software Engineering Agents for Repository-Level Code Maintenance, Debugging, and Issue Resolution
Keywords:
AI software development, code generation, AIAbstract
Large language models initially transformed software development through code completion and natural-language-to-code generation, but by 2025 research increasingly focused on autonomous software-engineering systems capable of navigating complete repositories, interpreting issue reports, executing commands, modifying multiple files, running tests, and iteratively correcting failures. This review examines the progression from pretrained code models to repository-level software-engineering agents through 2025. CodeBERT, GraphCodeBERT, CodeT5, and UniXcoder established transferable representations for source-code understanding and generation, while Codex and AlphaCode demonstrated substantial improvements in program synthesis. Repository-level benchmarks subsequently exposed limitations that were obscured by isolated programming tasks. RepoBench evaluated cross-file retrieval and completion, while SWE-bench formulated real GitHub issue resolution as an executable benchmark. SWE-agent showed that carefully designed agent-computer interfaces could substantially affect autonomous repair performance, whereas AutoCodeRover combined large language models with structured code-search mechanisms. Agentless demonstrated that comparatively simple localization, repair, and validation pipelines could remain competitive with more complex autonomous systems. OpenHands provided an open platform in which agents could write code, operate terminals, and browse the web. In 2025, SWE-Lancer expanded evaluation to real freelance engineering tasks with monetary value, while SWE-smith addressed the scarcity of scalable training environments for software-engineering agents. The review compares issue localization, repository retrieval, planning, patch generation, execution feedback, test-driven validation, sandboxing, cost, and benchmark design. Major challenges include long-horizon failure, incomplete specifications, flaky tests, unsafe command execution, dependency management, environment reproduction, benchmark contamination, and the gap between benchmark resolution rates and dependable professional software engineering.
References
Hindle A, Barr ET, Su Z, Gabel M, Devanbu P. On the naturalness of software. In: Proceedings of the 34th International Conference on Software Engineering. 2012. p. 837-847.
Allamanis M, Barr ET, Devanbu P, Sutton C. A survey of machine learning for big code and naturalness. ACM Comput Surv. 2018;51(4):81.
Chen M, Tworek J, Jun H, Yuan Q, Pinto HP, Kaplan J, et al. Evaluating large language models trained on code. arXiv. 2021;2107.03374.
Li Y, Choi D, Chung J, Kushman N, Schrittwieser J, Leblond R, et al. Competition-level code generation with AlphaCode. Science. 2022;378(6624):1092-1097.
Feng Z, Guo D, Tang D, Duan N, Feng X, Gong M, et al. CodeBERT: a pre-trained model for programming and natural languages. In: Findings of the Association for Computational Linguistics: EMNLP 2020. 2020. p. 1536-1547.
Guo D, Ren S, Lu S, Feng Z, Tang D, Liu S, et al. GraphCodeBERT: pre-training code representations with data flow. In: International Conference on Learning Representations. 2021.
Wang Y, Wang W, Joty S, Hoi SCH. CodeT5: identifier-aware unified pre-trained encoder-decoder models for code understanding and generation. In: Proceedings of EMNLP. 2021. p. 8696-8708.
Guo D, Lu S, Duan N, Wang Y, Zhou M, Yin J. UniXcoder: unified cross-modal pre-training for code representation. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics. 2022. p. 7212-7225.
Liu T, Xu C, McAuley J. RepoBench: benchmarking repository-level code auto-completion systems. In: International Conference on Learning Representations. 2024.
Jimenez CE, Yang J, Wettig A, Yao S, Pei K, Press O, Narasimhan K. SWE-bench: can language models resolve real-world GitHub issues? In: International Conference on Learning Representations. 2024.
Yang J, Jimenez CE, Wettig A, Lieret K, Yao S, Narasimhan K, Press O. SWE-agent: agent-computer interfaces enable automated software engineering. Adv Neural Inf Process Syst. 2024;37.
Zhang Y, Ruan H, Fan Z, Roychoudhury A. AutoCodeRover: autonomous program improvement. In: Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis. 2024.
Xia CS, Deng Y, Dunn S, Zhang L. Agentless: demystifying LLM-based software engineering agents. arXiv. 2024;2407.01489.
Wang X, Li B, Song Y, Xu FF, Tang X, Zhu M, et al. OpenHands: an open platform for AI software developers as generalist agents. In: International Conference on Learning Representations. 2025.
Miserendino S, Wang M, Patwardhan T, Heidecke J. SWE-Lancer: can frontier LLMs earn $1 million from real-world freelance software engineering? 2025.
Yang J, Lieret K, Jimenez CE, Wettig A, Khandpur K, Zhang Y, et al. SWE-smith: scaling data for software engineering agents. 2025.
Yao S, Zhao J, Yu D, Du N, Shafran I, Narasimhan K, Cao Y. ReAct: synergizing reasoning and acting in language models. In: International Conference on Learning Representations. 2023.
Schick T, Dwivedi-Yu J, Dessì R, Raileanu R, Lomeli M, Hambro E, et al. Toolformer: language models can teach themselves to use tools. Adv Neural Inf Process Syst. 2023;36.
Hong S, Zheng X, Chen J, Cheng Y, Zhang C, Wang Z, et al. MetaGPT: meta programming for a multi-agent collaborative framework. In: International Conference on Learning Representations. 2024.
Qian C, Liu W, Liu H, Chen N, Dang Y, Li J, et al. ChatDev: communicative agents for software development. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. 2024.
