Constraint Decay: The Fragility of LLM Agents in Back End Code Generation

LLM agents demonstrate strong performance in autonomous code generation but struggle when required to meet strict structural constraints needed for production-grade software, a phenomenon researchers term "constraint decay." The study found that agent performance declined significantly as structural requirements accumulated—with capable configurations losing 30 points on average in assertion pass rates—and identified data-layer defects as the leading cause of failures, particularly in convention-heavy frameworks like Django and FastAPI compared to minimal frameworks like Flask.