Robots now design their own brains

🔥 What happened Stanford researchers dropped RoboArchitect: a coding model that watches robot rollouts, diagnoses failures, and rewrites the agent harness around a frozen policy — no weight updates. Success jumps from 9.7% to 43.3% on unseen scenes. 💡 Why it matters That's a 4.5x gain from code alone. The evolved harness even beats finetuned specialists on out-of-distribution tasks. The bottleneck just moved from training to harness design — and that's automatable. ⚡ Our take If you're still burning GPUs on finetuning instead of optimizing the harness, you're lighting money on fire. This is software eating robotics.