AI-POWERED CODE TRANSLATION ANALYSIS

Scale-Dependent Benefits of Agentic AI

The effectiveness of the MCP framework varies significantly with model scale:

• 8B Models: MCP provides a crucial lift to pipeline progression (e.g., +14 pp function-level syntax pass) but sees a modest final success improvement due to limited semantic modeling capacity.
• 30B Models: This scale shows the greatest overall uplift. MCP's selective tool use and context hygiene are highly complementary, enabling the model to convert many failing candidates into successful simulations.
• 235B Models: At this high capacity, the baseline flow is already very competent. MCP offers less headroom, providing only a small lift. Interestingly, a naive RAG variant achieves the highest final success, suggesting larger models can filter non-helpful tokens and use naive retrieval as few-shot priming rather than distraction.

The Power of Conditional Tool Use

A central finding is that how auxiliary information is introduced matters as much as what information is available. Naively appending RAG outputs to every repair prompt (Non-MCP+RAG) is detrimental for smaller and mid-sized models. It causes context clutter, introduces architectural mismatches, and truncates precise compiler errors. The agentic framework's ability to selectively deploy tools based on the current context (e.g., only invoking RAG if a local fix fails) prevents these pitfalls, maintaining a compact, high-signal context crucial for less capable models.