Declarative Design Alone Won't Cut It: Study Shows Convention Alignment Drives AI-Friendly Code Frameworks

A team of computer scientists has challenged a widely held assumption about AI-assisted software development: that adopting a declarative design philosophy automatically makes code frameworks easier for AI assistants to work with.

In a new research paper submitted to arXiv's Computation and Language track, the authors introduce AI-assistability ($\mathcal{AI}$), a composite measure combining structural alignment with functional correctness to evaluate how amenable a framework is to AI-assisted code generation. Their findings suggest that the path to making frameworks more usable for AI isn't simply about elegant abstraction layers—it's about sticking to conventions that appear frequently in training data.

The Study Setup#

The researchers tested their metric using a novel benchmark called DDL2PropBank, which challenges AI systems to map database schemas to semantic role annotations. They then implemented the same agent logic across ten different multi-agent frameworks, allowing for direct comparison of how well language models could generate correct, framework-specific code in each one.

The setup revealed a striking pattern: Agno, a framework known for its single canonical pattern and convention-aligned API, achieved the highest AI-assistability score of 0.55. Meanwhile, DSPy—widely recognized as the most declaratively designed framework—scored just 0.07, largely because its novel abstractions don't appear frequently enough in the training data used by contemporary language models.

Convention Alignment Emerges as the Key Driver#

The analysis uncovered a correlation coefficient of 0.576 between structural alignment and functional correctness, indicating that adherence to established patterns is the strongest predictor of AI-assistability. This finding reframes the conversation around framework design: rather than competing on formal elegance or architectural purity, frameworks that stay close to familiar conventions are more likely to be effectively extended by AI coding assistants.

The implication is clear—when designing for AI assistance, embracing widely recognized patterns from the broader developer ecosystem may trump novel abstractions, regardless of how theoretically superior the latter might be.

All code, benchmarks, and evaluation artifacts have been made publicly available by the researchers to support further investigation in this emerging area of AI-assisted framework design.