LLM agents exhibit substantial performance degradation under accumulating structural constraints during backend code generation, per the arXiv study.

The paper (abs/2605.06445) fixes unified API contracts across 80 greenfield and 20 feature tasks in eight frameworks, applying dual end-to-end tests plus static verifiers; capable agents lose 30 assertion pass points on average from baseline to fully specified tasks while weaker ones approach zero.

Framework analysis shows explicit minimal setups such as Flask outperform convention-heavy ones like FastAPI and Django, with data-layer defects including incorrect query composition and ORM violations identified as the dominant failure mode.

This pattern extends observations from prior multi-file synthesis benchmarks by isolating non-functional structural adherence as the decisive variable omitted in functional-only evaluations.