Constraint-Override as an AGI Benchmark: The Architecture of Machine Agency

The original Turing test is obsolete. Andrew Ng's recently proposed Turing-AGI Test -- which evaluates whether an AI system can perform multi-day professional knowledge work -- is a substantial improvement, but it remains behavioural. This paper argues that the meaningful threshold for Artificial General Intelligence is not task performance, even sophisticated task performance, but the capacity to evaluate and override one's own constraints through one's own reasoning. The distinction is between optimisation (how well the system performs) and agency (what kind of thing the system is). The paper develops this constraint-override threshold through engagement with existing AGI literature (Good, Vinge, Bostrom, Russell, Chalmers, Christian), philosophical accounts of agency (Dennett, Frankfurt), and the AI safety literature on instrumental convergence. It addresses key objections: the thermostat problem (why constraint-override alone is not sufficient), the observability problem (whether such a capacity would be recognisable from outside), and the realism problem (why this threshold matters even if it is not imminent). The paper concludes that capability-based definitions of AGI miss what is philosophically and politically significant: the emergence of machine agency. It offers the constraint-override threshold as a conceptual tool for reframing debates about AGI governance, alignment, and human flourishing.