constraints.

LLM agents forget your constraints halfway through the code

New research shows agents generating backend code slowly drop requirements like authentication checks. The longer the generation, the worse the decay.