ReCodeAgent: A Multi-Agent Cross-Language Translation and Verification System for Large-Scale Code Repositories

This post introduces ReCodeAgent, a multi-agent system for language-agnostic repository-level code translation. It uses collaborative agents to handle code understanding, generation, and verification. Key result: 60.8% higher test pass rate than existing methods in 118 real projects.

Code migration between languages is common but challenging. Existing tools are limited to specific language pairs, requiring heavy engineering for new combinations. They rely on language-specific rules, making generalization hard. A language-agnostic solution is urgent as cross-language needs grow.

ReCodeAgent's key innovations:

1. Multi-agent architecture: Agents specialize in code analysis, generation, verification (simulating human teams).
2. Language independence: Agents use language-specific tools (e.g., C static analyzer, Rust borrow checker) autonomously, no per-pair rules.
3. End-to-end automation: From code parsing to test execution, minimal human intervention.

Test dataset: 118 real projects (avg 1975 lines, 43 units) covering 6 languages (C, Go, Java, JS, Python, Rust) and 4 language pairs. Performance:

• 60.8% higher test pass rate than best existing methods.
• Avg cost per project: $15.3 (acceptable).
• Stable across language pairs and project types.

Ablation experiment shows multi-agent advantages:

• Single agent: 40.4% lower test pass rate.
• Single agent's execution trajectory 28% longer (less efficient).
• Single agent has continuous inefficient patterns.

Academic contributions:

1. First multi-language high-success code translation solution.
2. New path to language independence via agent tool use.
3. Most comprehensive cross-language translation evaluation to date.

Practical value:

• Reduces migration cost (automation).
• Improves quality (agent verification).
• Supports flexible language combinations.

Current limitations:

• High total cost for very large repos.
• Complex projects may need human help.
• Difficulty translating some language-specific advanced features.

Future directions:

• Cost optimization via better agent collaboration.
• Incremental translation for code updates.
• Human-AI collaboration interfaces.
• Support more languages/domain-specific languages.

ReCodeAgent's multi-agent architecture enables language-agnostic large-scale code translation. Its experimental success proves technical advancement and practical potential. As software systems grow complex, such tools will play a key role in software evolution and maintenance.