This working paper examines a shift in AI-assisted systems from prompt-level interaction toward externalized state architectures. Prompt engineering changes individual responses, but increasingly important forms of adaptation occur through durable external objects such as skill files, workflow loops, memory systems, registries, and validation structures. The paper uses SkillOpt as its primary technical anchor. SkillOpt optimizes a natural-language skill document while keeping the underlying model fixed, providing evidence that procedural behavior can be improved through trainable external state. The paper compares this approach with Karpathy-style AutoResearch workflows and automated-research systems such as AI Scientist. Drawing on cybernetics, the paper argues that these systems are best understood as feedback-regulated architectures whose behavior depends not only on model weights but also on the state they preserve, update, and govern. The Reflexive Laboratory is used as a case of research-scale state engineering, where working-state derivation, admissibility review, canonicality rules, and artifact-integrity validation become part of the capability architecture. The paper's central claim is that the design problem after prompt engineering is increasingly a problem of governing externalized state.
Peter Bell (Mon,) studied this question.