DOI: 10.17587/prin.17.535-541

Controlled Improvement of Software Agents based on Large Language Models using a Fitness Function

F. G. Zhuchkov¹, Doctoral Cand., philipgermanovich@yandex.ru, E. R. Bryukhanova^{2, 3}, Cand. Sc. (Eng.), Associate Professor, Senior Systems Analyst, evgbryuhanova@gmail.com, O. V. Kolosova¹, D. Sc. (Eng.), Professor of the Higher School of Technological Entrepreneurship, okolosova@spbstu.ru, A. M. Khakhina¹, D. Sc. (Eng.), Professor of the Higher School of Software Engineering, hahina_am@spbstu.ru, O. A. Antamoshkin³, D. Sc. (Eng.), Head of the Department of Software Engineering, oleslav@mail.ru
¹ Peter the Great St. Petersburg Polytechnic University, Saint Petersburg, Russian Federation,
² RUT CODE JSC, Moscow, Russian Federation,
³ Siberian Federal University, Krasnoyarsk, Russian Federation

Corresponding author: Oleslav A. Antamoshkin, D. Sc. (Eng.), Head of the Department of Software Engineering, Siberian Federal University, 660041, Krasnoyarsk, Russian Federation, E-mail: oleslav@mail.ru

The paper addresses the problem of developing software agents based on large language models capable of automatically performing individual operations related to software code modification, result verification, and accumulation of experience from previous actions. An approach to the controlled improvement of software systems is proposed, in which an agent generates hypotheses of code modifications, applies changes to the codebase, executes a testing pipeline, and evaluates the obtained result using a quality assessment function. Process controllability is ensured through fixed experimental conditions, automated verification of modifications, and comparison of different agent behavior strategies. For experimental evaluation, the RSI-Bench framework was developed to compare strategies of software agent behavior in code modification tasks. The primary evaluation metric was the success rate of automated test execution. Additionally, computational cost and the number of iterations required to obtain a result were recorded. The conducted experiments demonstrated that solution quality depends not only on the selected agent behavior strategy, but also on the underlying language model, the specific agent implementation, and stochastic variability between runs. It was found that self-modification strategies implemented at the level of textual instructions provide a limited and unstable effect. The obtained results may be used in the development of automated software engineering support systems and approaches for assessing the quality of software code modifications.

Keywords: software engineering, large language models, software agents, software code generation, self-modification, fitness function, search-based optimization, software quality assessment

pp. 535—541

For citation:
Zhuchkov F. G., Bryukhanova E. R., Kolosova O. V., Khakhina A. M., Antamoshkin O. A. Controlled Improvement of Software Agents based on Large Language Models using a Fitness Function, Programmnaya Ingeneria, 2026, vol. 17, no. 10, pp. 535—541. DOI: 10.17587/prin.535-541 (in Russian).

References:

1. Yang J., Jimenez C. E., Wettig A. et al. SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering. arXiv preprint, 2024, arXiv:2405.15793. DOI: 10.48550/arXiv.2405.15793.
2. Zhang Y., Ruan H., Fan Zh., Roychoudhury A. Auto-CodeRover: Autonomous Program Improvement, Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA '24), New York, ACM, 2024, рр. 1592—1604. DOI: 10.1145/3650212.3680384.
3. Shinn N., Cassano F., Gopinath A., Narasimhan K., Yao Sh. Reflexion: Language Agents with Verbal Reinforcement Learning, Advances in Neural Information Processing Systems, 2023, vol. 36, рр. 8634—8652. DOI: 10.52202/075280-0377.
4. Besta M., Blach N., Kubicek A. et al. Graph of Thoughts: Solving Elaborate Problems with Large Language Models, Proceedings of the AAAI Conference on Artificial Intelligence, 2024, vol. 38, no. 16, pp. 17682—17690. DOI: 10.1609/aaai.v38i16.29720.
5. Zhou A., Yan K., Shlapentokh-Rothman M. et al. Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models, Proceedings of Machine Learning Research, 2024, vol. 235, pp. 62138—62160.
6. Harman M. The Current State and Future of Search-Based Software Engineering, Future of Software Engineering (FOSE 2007). Washington, IEEE Computer Society, 2007, pp. 342—357. DOI: 10.1109/FOSE.2007.29.
7. Harman M., Mansouri S. A., Zhang Y. Search-Based Software Engineering: Trends, Techniques and Applications, ACM Computing Surveys, 2012, vol. 45, no. 1, article 11. DOI: 10.1145/2379776.2379787.
8. McMinn P. Search-Based Software Test Data Generation: A Survey. Software Testing, Verification and Reliability, 2004, vol. 14, no. 2, pp. 105—156. DOI: 10.1002/stvr.294.
9. Antamoshkin O. A. Software Engineering: Theory and Practice. Textbook. Krasnoyarsk, Siberian Federal University, 2012, 247 p. (in Russian).
10. Antamoshkin O. A., Antamoshkina O. A., Smirnov N. A. Multi-agent automation system for monitoring, forecasting and managing emergency situations, IOP Conference Series: Materials Science and Engineering, 2016, vol. 122, article 012003. DOI: 10.1088/1757-899X/122/1/012003.
11. Sartaj H., Ali Sh., Arcaini P., Arcuri A. Search-Based Software Engineering and AI Foundation Models: Current Landscape and Future Roadmap, arXiv preprint, 2025, arXiv:2505.19625. DOI: 10.48550/arXiv.2505.19625.