HyperAgent: Generalist Software Engineering Agents to Solve Coding Tasks at Scale

  • 2024-11-05 17:22:10
  • Huy Nhat Phan, Tien N. Nguyen, Phong X. Nguyen, Nghi D. Q. Bui
  • 0

Abstract

Large Language Models (LLMs) have revolutionized software engineering (SE),showcasing remarkable proficiency in various coding tasks. Despite recentadvancements that have enabled the creation of autonomous software agentsutilizing LLMs for end-to-end development tasks, these systems are typicallydesigned for specific SE functions. We introduce HyperAgent, an innovativegeneralist multi-agent system designed to tackle a wide range of SE tasksacross different programming languages by mimicking the workflows of humandevelopers. HyperAgent features four specialized agents-Planner, Navigator,Code Editor, and Executor-capable of handling the entire lifecycle of SE tasks,from initial planning to final verification. HyperAgent sets new benchmarks indiverse SE tasks, including GitHub issue resolution on the renowned SWE-Benchbenchmark, outperforming robust baselines. Furthermore, HyperAgent demonstratesexceptional performance in repository-level code generation (RepoExec) andfault localization and program repair (Defects4J), often surpassingstate-of-the-art baselines.

 

Quick Read (beta)

loading the full paper ...