According to CoinMeta, well-known AI reviewers of ARC-AGI, such as Fran, ois, and Chollet, state that today's high-performance AI and agent are essentially complete systems composed of "model + program." The large model is responsible for understanding and judgment, while the outer program is in charge of controlling the process, invoking tools, and running code. He described this neural-symbolic architecture as a "symbol sandwich," believing that large models will continue to become stronger, but the outer programs will also become increasingly important. Ultimately, complex tasks will require the cooperation of both. In addition, the founder of RAFT, Richard Qian, proposed that the stronger the model, the agent and harness may become thicker; many fixed processes can be directly handed over to the model, and in the end, there might only be one bash tool left.