ARC-AGI Creator Predicts Future AI Competition Will Focus on Full Agent Systems

According to Beating Monitor, François Chollet, creator of the renowned AI evaluation ARC-AGI, stated that today’s high-performance AI agents are essentially complete systems composed of “model + program.” The large model handles understanding and judgment, while the outer program manages workflow, calls tools, and executes code. Chollet describes this neural-symbolic architecture as a “symbolic sandwich”: the program drives the neural network, and the neural network, in turn, calls upon the program. He believes that while large models will continue to grow stronger, the outer program will become increasingly important as well—complex tasks will ultimately require collaboration between both components. Furthermore, Chollet predicts that the outer programs, currently primarily hand-written by engineers, may one day be learned and evolved by AI itself. In the future, competition among AIs may no longer be about models alone, but about the entire agent system.