From Issue to Pull Request: A Multi-Agent Orchestration Framework for End-to-End Software Engineering Automation

Large Language Models have demonstrated strong capabilities in isolated software engineering tasks, but automating the full development lifecycle in enterprise codebases remains an open challenge. We present an agentic pipeline framework that orchestrates multiple specialized LLM agents through a graph-based state machine to automate end-to-end software engineering workflows — from issue analysis to production-ready pull request. The framework employs self-healing feedback loops that automatically reroute execution when quality gates fail, domain skill injection for codebases-specific knowledge without model fine-tuning, and the Model Context Protocol (MCP) for IDE-embedded execution. Evaluation on 127 production workflows in a large-scale Java monorepo demonstrates a 67% end-to-end success rate, with self-healing loops improving the all-gates pass rate from 22% to 67% — a 45 percentage-point improvement. Developer effort reduction ranges from 60% to 73% depending on task complexity.