AI Agents Build 3D Scenes – Blind to Errors

🔥 What happened Researchers at the University of Maryland and AWS unveiled LEGO-Anything: a coding agent writes an executable Blender program from a single photo. The best tested agent, GPT-6 Astra, hits only 53.4% accuracy on indoor and 39.6% on outdoor scenes. 💡 Why it matters The agent can't judge its own progress—when picking which of two versions is better, it performs at chance level. The LEGO-Plugin boosts weak models by up to 62.7%, while the top model gains just 2 percentage points. ⚡ Our take If you build an agent that grades its own homework, you get an agent that lies to itself. Measure everything, trust nothing.