Machine Studying: A System-Level Reframing of Continual Adaptation from Declarative Corpora

Humans study in various ways, like reading textbooks, working through examples, or thinking out loud. Underneath these activities is the capacity to convert preparation into expertise, and we argue that this belongs at the center of how agentic AI systems should adapt. We call the machine analogue of it $\textit{machine studying}$: the work that an agentic system performs on itself, given a corpus and no preview of the downstream tasks, that pays off at test time. Potential approaches for machine studying are open-ended, ranging across gradient descent, harness design, context optimization, and combinations thereof. The payoff from studying manifests as fewer mistakes or faster answers. We instantiate the framework in an initial setting in which agents must teach themselves the concepts and usage of an extensive software library from nothing but a corpus of its internal code, with conceptual and coding tasks and study procedures that act on weights, context, and harness. We offer this as the new frame through which continual adaptation should be studied.