Should we define AI agents as a spectrum rather than a category?

OpenAI's 2023 white paper on governing agentic AI systems defines an AI system's "agenticness" not as a yes/no label but as "the degree to which a system can adaptably achieve complex goals in complex environments with limited direct supervision," scored along four components. Goal complexity asks how hard the goal would be for a human and how wide a range of goals the system can pursue — an AI that answers both legal and programming questions has higher goal complexity than a classifier that only sorts the two. Environmental complexity asks how cross-domain, multi-stakeholder, long-horizon, or multi-tool the operating environment is — an AI that plays any board game well has higher environmental complexity than one that only plays chess. Adaptability asks how well the system handles novel or unexpected circumstances — human customer-service representatives have higher adaptability than rule-based bots. Independent execution asks how reliably the system achieves goals with limited human intervention — a level-3 autonomous car has higher independent execution than a traditional car.

The paper's reasoning for using four axes instead of a category is explicit: "there is no clear line along which to draw a binary distinction between 'agents' and current AI systems like GPT-4," so agenticness is better treated "as a property rather than a category/classification." It also narrows what the term covers by exclusion: agenticness is explicitly separated from consciousness, moral patienthood, and self-motivation, and from anthropomorphism, and the paper notes it is "not tied to physicality" — a software assistant can be more agentic by this definition than most robots. The four axes let a system score high on one dimension, such as independent execution, while scoring low on another, such as environmental complexity, which a single agent/non-agent line cannot represent.

The move away from a binary line is shared with Does machine agency exist on a spectrum rather than binary?, but the two taxonomies classify differently: Rammert's levels — passive, semi-active, reactive, proactive, cooperative — describe a system's behavior mode as a single rung, while OpenAI's four axes score a system along independent dimensions that do not collapse into one rung. The later, more empirical claim in Does a single benchmark score actually predict agent readiness? about 2026 benchmarks is a direct descendant of this 2023 conceptual argument — both insist capability is multi-dimensional, but the 2026 work operationalizes specific measurable axes (task success, privacy compliance, long-horizon retention) where this paper offers conceptual ones. Independent execution, one of the four axes here, is the same autonomy dimension that Does AI risk increase with the autonomy we give it? treats as the primary driver of risk — this paper supplies the vocabulary for naming that dimension precisely.

The definition is conceptual, offered to frame the rest of the paper's governance discussion, not a measurement instrument: the excerpt gives illustrative examples — chess versus any board game, rule-based bots versus human representatives — rather than a method for scoring a real system on any axis, and it does not say how the four axes combine into an overall judgment of whether a system is "agentic enough" to warrant particular safety practices. It is useful for naming what varies between systems, not for certifying any one system's readiness.