The Body of Software

A model can produce a plan and even call a tool without having a dependable means of acting in its environment. I came to this question while developing Loki, an Android interface intended to be occupied by different AI backends while the device, its permissions, schedules and practical capabilities remain separate from them. This arrangement asks us to distinguish the intelligence selecting an action from the body through which that action becomes possible. Existing work on embodied cognition and computer-use agents explains much of the perception–action relationship, while security engineering already provides principles for mediation and accountability. I examine where these ideas meet, and propose a Causal Embodiment Contract for assessing an operational software body. Its central test is a model substitution: if the controller changes, which properties of the body remain stable, which useful abilities change, and can an improvement in the controller reveal a permission boundary that was never complete? The paper develops a comparative model, source-attributed evidence from Loki, non-empirical counterexamples and a protocol for future device tests. It does not report a controlled Android experiment. Its contribution is a way to investigate the body as a continuing unit of engineering, rather than judge agency solely by the model or the success of an individual tool call.