Agents optimize their own stack

Agentic AI systems are being used not only to write application code but to generate, test and optimize low‑level infrastructure (kernels, TPU code, device drivers). These closed‑loop agents produce verified traces that can be fed back as high‑quality synthetic training data, accelerating both model capability and hardware/software co‑optimization. — If agents routinely optimize the compute stack, control over AI capability will shift from raw chip supply or data scale to who operates closed‑loop optimization pipelines, with implications for industrial policy, energy use, security, and market concentration.