The Evolution of the Agent Harness
Around Christmas 2025, AI engineers observed that agents began to function effectively. The authors believe that this is the result of the intersection of model capabilities and the maturity of the agent framework at a specific point in time. Prior to this, ChatGPT in November 2022 only had word prediction and human feedback reinforcement learning capabilities; it lacked tools, search, and reasoning functions, and was likened to a “brain in a tank.” The agent framework is a system that enables large language models to overcome the limitations of training data and interact with real digital spaces, including components such as environments, tools, context, and guards. The authors divided the evolution of the framework into three stages: ReAct technology in October 2022 served as a “paper framework,” defining an intelligent loop outside of weights; Toolformer proposed by Meta in February 2023 suggested that tool use could be internalized during training; and AutoGPT/BabyAGI in the spring of 2023 made the framework’s capabilities exceed those of the model, giving the model excessive autonomy.