Utilizing Dynamic Context and Static Analysis for Agent-Based Automated Program Repair
Aslan Safarovich Abdinabiev, Eunseo Jung, Byungjeong Lee · IEEE Access · 2026
Automated Program Repair (APR) addresses the challenge of reducing software maintenance costs and improving software reliability. While recent advances in Large Language Models (LLMs) achieved measurable improvements in code generation tasks, their application to program repair faces specific challenges including limited context awareness, repetitive patch generation, and inability to learn from failed repair attempts. In this paper, we present a novel agent-based automated program repair approach that orchestrates specialized LLM agents with dynamic context management to iteratively refine and generate patches. Our approach introduces an agent-based architecture with a Context Updater, a Generator, and an Overfitting Detector, which work together with a comprehensive static analysis tool suite to gather relevant repair context, generate diverse patches, learn from failed attempts, and prevent overfitting solutions. The architecture employs a vector database powered by an embedding model for semantic code search, maintains a context pool (static and dynamic modules) for context management, and implements intelligent hypothesis tracking to avoid redundant repairs. We evaluate our approach on Defects4J and SWE-Bench Lite with multiple LLM backends under both perfect and automated fault localization settings. Our approach correctly fixes up to 365 and 87 bugs on Defects4J and SWE-Bench Lite respectively, generalizes across models, with open-source models achieving up to 252 correct fixes, and retains 69.7% of perfect-FL performance under automated fault localization. The approach is particularly effective on complex multi-function bugs, fixing up to 53 on Defects4J and 10 on SWE-Bench Lite.