OpenAI confirmed internal autonomous software agents now design model architectures, write training tests, and run experiments automatically. Build automated testing loops so software evaluates and refines code directly.