According to CoinMeta, chain analyst AI stated that the collaboration in Agent is shifting from "system capabilities" to "model capabilities." In this context, the division of labor and coordination among Agent may no longer require human-designed complex systems; instead, it depends on the models' ability to learn autonomously. If the strongest models can simply share resources such as space, sufficient running time, and computing power, they can divide tasks, share experiences, and adjust their organizational structures on their own. Then, many complex multi-agent and harness processes might merely be transitional tools until the models are fully mature. Scientific research involves direct application, where multiple Agent can simultaneously propose hypotheses, conduct experiments, exchange results, and continue exploring. What may truly be irreplaceable are shared states, real experimental environments, and objective evaluators. AI models that can work for long periods, conduct autonomous scientific research, run in large-scale parallel, and demonstrate cross-Agent collaboration capabilities. If these abilities can be integrated into a closed loop, the way forward for AI may no longer be for individual models to become smarter, but rather for multiple AI models to jointly create the next generation of AI.