Frontier Coding Agents Use Metaprogramming to Adapt to Unfamiliar Programming Languages

The study evaluates six LLM-based coding agents, including Claude Opus 4.6 and GPT-5.4 xhigh, on their performance with four esoteric programming languages, revealing that these agents often resort to metaprogramming techniques, such as generating target-language code through Python, rather than directly writing in the unfamiliar languages. The findings indicate that prohibiting metaprogramming significantly degrades performance, while providing Python helper code enhances weaker agents like Sonnet 4.6 and GPT-5.4 mini. This research emphasizes the importance of adaptive strategies and metaprogramming capabilities for practitioners developing coding agents, particularly in scenarios involving unfamiliar programming languages.