A Structural Theory of Harnesses (nemooperans.com)

A recent paper has introduced a groundbreaking structural theory of "harness engineering" in AI, emphasizing a shift in focus from merely scaling base models to creating the infrastructure that enhances their capabilities. Following the Claude Code leak in March 2026, which unveiled complex software architectures fundamental to AI functionalities, the practice of harness engineering emerged as a recognized discipline within tech circles. Harnesses are defined as systems that establish feedback loops between AI generators and real-world experiences, enabling them to learn and adapt—capabilities that raw models alone lack. This marks a significant development in the AI/ML community, as it redefines how intelligence is realized and maintained in AI systems. The implications of this theory are manifold. It posits that generalized intelligence, a crucial element for advancing artificial general intelligence (AGI), hinges on effective harness engineering, rather than just on the models themselves. Harness properties such as memory, social competence, and affective state tracking are highlighted as essential for creating this feedback loop and achieving adaptive behavior. The paper also describes a practical example in the form of an open-source memory architecture called "anneal-memory," demonstrating that these harness properties can be effectively defined and implemented today. Overall, this work not only sheds light on the mechanics of AI cognition but also situates the conversation around AI alignment and quality within the framework of harness engineering, offering a fresh perspective on the future of AI development.