Survey Maps LLM-Based Agents for Automating Software Bug Fixes and Issue Resolution

Researchers have completed a comprehensive examination of how large language models (LLMs) are being deployed as autonomous agents to resolve software bugs and issues in real repositories. The work, submitted to arXiv in late 2025 and revised in August 2026, synthesizes insights from 242 recent studies at the intersection of AI and software engineering.

Why This Matters#

Maintaining software systems is labor-intensive and error-prone. Software issue resolution—the process of identifying and fixing problems described in natural language—is central to keeping codebases healthy. Traditional approaches require human developers to manually diagnose and implement fixes, which consumes significant time and expertise. LLM-based agents offer the potential to automate substantial portions of this work, promising improvements in both efficiency and quality while reducing the burden on engineering teams.

The Challenge of Real-World Problem-Solving#

Fixing actual software issues is far more complex than single-step tasks. It demands long-horizon reasoning—the ability to plan multiple steps ahead—iterative exploration through potential solutions, and feedback-driven decision-making that adapts to new information. These capabilities go beyond what conventional LLM approaches (which process input and generate output in one pass) can handle. Agentic systems, by contrast, loop continuously: they perceive the environment, reason about next steps, take action, observe results, and refine their strategy accordingly.

Survey Scope and Structure#

The survey organizes its findings across three primary dimensions:

• Benchmarks — standardized datasets and evaluation frameworks for measuring agent performance on software tasks
• Techniques — methods and architectural approaches for building and improving agentic systems
• Empirical studies — real-world results and case analyses showing how these agents perform in practice

Emerging Training Paradigm#

One key insight from the literature is the rising importance of reinforcement learning (RL) as a training strategy. Rather than relying solely on supervised learning from labeled examples, RL allows agentic systems to learn from trial and error, receiving feedback signals that guide them toward better solutions. This approach aligns naturally with the iterative, exploratory nature of software debugging.