AI doesn’t fail because models are weak.

It fails because nobody can explain what just happened.

New research shows agent systems struggle with reproducibility and hidden state — not intelligence, but coherence.