https://arxiv.org/abs/2604.14228 pretty simple, ai-trust philosophy interfaces / agent runtime / hooks&tools&policies / memory / sandbox warning: ai-assisted devs show 17% degradations in code understanding, so keep grounded, folks