In this post, researcher Lilian Weng examines how harness engineering, the systems that orchestrate model execution, tool use, and workflow logic, contributes to recursive self-improvement in AI agents. The article discusses design patterns such as workflow automation, file-based persistent memory, and parallel sub-agents, alongside optimization approaches ranging from context engineering to evolutionary search. Weng argues that the layer between a raw model and its real-world context is as important as the model’s underlying capability, and presents harness code as a way to expand the AI optimization design space beyond manual prompting.
