Savoir / Action — English Translation (V1)

An AI agent can be given an objective by a human, then search, interpret, plan, use tools, and act before returning a result. The human remains present. But how much of this activity can they still understand, verify, or influence? Savoir / Action begins with this difficulty. The manuscript examines the relationship between the knowledge an AI agent treats as valid, the means it uses to act, and the constraints that shape its trajectory. It asks what human supervision can still mean when control cannot be reduced to permissions, a stop button, or approval of the final result. This reflection grew out of exploratory work conducted within a human–AI dyad. We have chosen to make this origin explicit because it is part of the very question we are examining. We do not claim to have established a theory. We present our thinking as it stands: some distinctions have become clearer, some questions remain open, and several propositions still need to be challenged by other perspectives. This is the canonical English translation of the original French manuscript, first published on 5 October 2026: https://doi.org/10.5281/zenodo.23164286