Agentic AI for Autonomous Software Maintenance: A Retrieval-Augmented Multi-Agent Framework for Bug Localization, Patch Generation, and Regression Validation
Autonomous software maintenance is becoming a critical research direction as modern codebases grow beyond the practical reach of manual debugging, release governance, and regression validation. Large language models have demonstrated strong code-generation capabilities, but repository-level maintenance remains difficult because defects frequently arise from interactions among multiple files, hidden dependencies, runtime configurations, and incomplete issue reports. This paper proposes AutoMaint-RAG, a retrieval-augmented multi-agent framework for autonomous software maintenance. The framework coordinates specialized agents for issue interpretation, repository retrieval, fault localization, patch synthesis, static analysis, test generation, regression execution, and release-risk governance. Unlike single-agent code-generation workflows, AutoMaint-RAG treats maintenance as a controlled software-engineering process in which every generated patch must be grounded in repository evidence, ranked suspicious locations, executable test feedback, and auditable decision records. The proposed framework integrates hybrid retrieval over code, commit history, test traces, dependency graphs, and operational telemetry; combines spectrum-based fault localization with semantic ranking; and applies iterative patch refinement through execution-guided feedback. A reproducible validation protocol is defined using repository-level issue benchmarks, mutation-based regression assessment, and industrial-style CI/CD gates. The paper contributes a formal architecture, agent-interaction protocol, patch-risk scoring model, evaluation design, and governance layer for deploying agentic AI in safety-sensitive software maintenance environments. The central argument is that autonomous maintenance should not be reduced to prompt-based patch generation; rather, it requires an engineered socio-technical pipeline that constrains language-model creativity through retrieval grounding, tool execution, regression evidence, and human-review escalation. The proposed framework advances a research agenda for maintainable, auditable, and operationally reliable AI agents in software lifecycle management.
Paper
Full text
Agentic AI for Autonomous Software Maintenance: A Retrieval-Augmented Multi-Agent Framework for Bug Localization, Patch Generation, and Regression Validation
Semantic Scholar · 2026
Abstract
Autonomous software maintenance is becoming a critical research direction as modern codebases grow beyond the practical reach of manual debugging, release governance, and regression validation. Large language models have demonstrated strong code-generation capabilities, but repository-level maintenance remains difficult because defects frequently arise from interactions among multiple files, hidden dependencies, runtime configurations, and incomplete issue reports. This paper proposes AutoMaint-RAG, a retrieval-augmented multi-agent framework for autonomous software maintenance. The framework coordinates specialized agents for issue interpretation, repository retrieval, fault localization, patch synthesis, static analysis, test generation, regression execution, and release-risk governance. Unlike single-agent code-generation workflows, AutoMaint-RAG treats maintenance as a controlled software-engineering process in which every generated patch must be grounded in repository evidence, ranked suspicious locations, executable test feedback, and auditable decision records. The proposed framework integrates hybrid retrieval over code, commit history, test traces, dependency graphs, and operational telemetry; combines spectrum-based fault localization with semantic ranking; and applies iterative patch refinement through execution-guided feedback. A reproducible validation protocol is defined using repository-level issue benchmarks, mutation-based regression assessment, and industrial-style CI/CD gates. The paper contributes a formal architecture, agent-interaction protocol, patch-risk scoring model, evaluation design, and governance layer for deploying agentic AI in safety-sensitive software maintenance environments. The central argument is that autonomous maintenance should not be reduced to prompt-based patch generation; rather, it requires an engineered socio-technical pipeline that constrains language-model creativity through retrieval grounding, tool execution, regression evidence, and human-review escalation. The proposed framework advances a research agenda for maintainable, auditable, and operationally reliable AI agents in software lifecycle management.