PatchPilot presents a novel framework that combines the strengths of LLM-driven and human-designed workflows to create more reliable automated software patching systems.

• Achieves 90.5% relative performance of leading LLM agents while reducing costs by 83%
• Implements a three-phase patching pipeline (understanding, planning, implementation) that dramatically improves reliability
• Introduces specialized modules for context retrieval, patch verification, and debugging to handle complex software repair scenarios
• Delivers consistent performance across diverse software projects without requiring costly model fine-tuning

This research represents a significant advancement for Engineering teams by making automated bug fixing more practical and cost-effective for real-world deployment, potentially reducing development cycles and improving software quality.

PatchPilot: A Stable and Cost-Efficient Agentic Patching Framework