AI
Multi-Agent System Translates Code Across Programming Languages at Scale
Researchers introduce ReCodeAgent, an autonomous workflow that can convert entire repositories from one programming language to another with high accuracy across six different languages.
Automating the translation of code between programming languages has long been a challenge for software engineering. Most existing solutions work well for only a single language pair, requiring substantial effort to adapt them for new combinations. A new system called ReCodeAgent tackles this problem by using multiple AI agents working together to handle repositories regardless of source or target language.
How It Works#
ReCodeAgent operates as an autonomous multi-agent workflow. Users provide a software project written in one programming language and specify which language they want it translated into. The system then automatically translates the entire repository and validates the results using language-specific testing tools. Rather than building custom translation pipelines for each language pair, ReCodeAgent leverages AI agents capable of synthesizing code across many languages and autonomously executing the appropriate analysis and testing tools for each one.
Testing Results#
The system was evaluated on 118 real-world projects spanning six programming languages across four different translation pairs. On average, each project contained around 1,975 lines of code organized into 43 translation units. When compared against four alternative neuro-symbolic and agentic approaches, ReCodeAgent demonstrated superior translation accuracy. The system achieved a 60.8% improvement in test pass rate on ground-truth test cases, with an average operational cost of $15.3 per project.
Design Trade-Offs#
The researchers also examined how architectural choices affect performance. Switching from the multi-agent design to a single-agent approach resulted in significant performance degradation: test pass rates dropped by 40.4% on average, and the workflows became 28% longer while remaining inefficient. This analysis underscores the importance of the coordinated multi-agent approach to the system's effectiveness.
The work represents the first technique to achieve high translation success rates consistently across multiple programming languages, addressing a longstanding gap in code translation research.
