Autonomous Code Refactoring Agent: A Legacy Monolithic Architecture Modernization Solution Based on Multi-Agent Workflow

This article introduces a legacy monolithic architecture modernization solution based on multi-agent workflow, built using OpenCode, Aider, and Claude/GPT models. It enables autonomous scanning, understanding, refactoring, and validation of legacy code, ensuring refactoring quality through a closed-loop verification mechanism and significantly reducing technical debt and maintenance costs.

Originating from an enterprise resource planning (ERP) platform (PHP, MySQL, Apache stack) that has been in operation for over a decade, code structure degradation has occurred due to feature iterations, requirement changes, etc.: highly coupled business logic, undocumented dependencies, and lack of modularization. Team pain points: high investment in adding simple features, high regression risk during deployment; senior engineers spend 40% of their weekly time on code sorting, vulnerability identification, and PR reviews—manual supervision is expensive and error-prone.

The goal is to automate the tedious aspects of technical debt reduction and free the team to focus on high-value work. It adopts a multi-agent workflow architecture, combining OpenCode and Aider frameworks, driven by Claude/GPT models to address legacy system modernization challenges.

The core innovation is the separation of strategic and tactical layers:

• OpenCode Strategic Layer: Claude model (with a large context window) is responsible for global repository scanning, dependency mapping, and business logic understanding;
• Aider Tactical Layer: GPT model handles micro-refactoring, rapid iterative modifications, writing unit tests, and resolving syntax errors. Separation of concerns balances deep architectural reasoning and precise code generation.

The system operates in a four-stage cycle:

1. Ingestion and Static Analysis: Scan directories, perform deep semantic analysis to identify anti-patterns (e.g., large functions violating SOLID principles);
2. Strategy Formulation: Generate a refactoring blueprint, plan decoupling solutions, and ensure complete business logic;
3. Execution and Collaboration: Assign tasks to Aider, execute code rewriting (e.g., SQL migration to ORM), and retain business logic;
4. Closed-Loop Verification: Automatically generate unit tests, execute tests, and if failed, analyze the cause, generate patches for fixes until 100% pass rate is achieved.

After deployment, it changed team operations:

• Reduced review time, freeing up senior engineers;
• Lowered defect rate, with closed-loop verification automatically capturing and fixing regressions;
• Controllable technical debt, enabling continuous modernization;
• Generated documents and test cases serve as living documents.

Insights: Tool combination (complementing advantages of different models), verification mechanism (autonomous verification is key), progressive modernization (avoiding large-scale rewrite risks). Future: Extend AI agents to software engineering fields such as architecture design, performance optimization, and security auditing.