Should we judge AI systems as 'agent or not' — or ask how agentic they are, and in which specific ways?

This explores what changes when you stop asking 'is this an AI agent or not?' and start asking 'how agentic is it, and in which ways?' The clearest graded model in the collection comes from OpenAI, which scores systems on four separate axes: how complex the goals are, how complex the environment is, how well the system adapts, and how much it executes without supervision Should we define AI agents as a spectrum rather than a category?. The key move is that the axes are independent. A system can pursue simple goals with a lot of autonomy, or work on hard goals while a human approves each step. A binary definition would have to force both into the same box. A profile keeps the difference visible, and that difference is where most of the safety and design questions sit.

The binary debate hasn't gone away, though. It has moved to a more basic question: are these systems agents at all? Farrell and colleagues argue that large models are better understood as cultural technologies, closer to markets or bureaucracies Are large AI models actually intelligent agents or social technologies?. On their view, a model compresses and reorganizes what humans already know without having any intentions of its own. Classical multi-agent systems research sets the bar from the other side. Its agents have explicit, inspectable beliefs, desires and intentions, which LLM agents lack. Their 'agency' is implicit and pattern-based Can classical agent architectures fix agentic AI accountability?. Both critiques are binary in spirit, and both point to something a graded score can hide. A system can rank high on independent execution and still have nothing inside it that you could call a goal.

The surprising thread is that agency may not belong to the model at all. Several notes place it in the surrounding system. One argues that reliable agents get their reliability by moving memory, skills and interaction protocols into an outside 'harness' rather than leaving them to the model Where does agent reliability actually come from?. Another argues that real tasks outgrow any single agent loop and need several specialized agents working together Do single agents always hit organizational limits?. In multi-agent setups, even accountability tags only change outcomes when the workflow actually routes work through the agent that holds them Does accountability placement actually shape multi-agent outcomes?. On this reading, 'how agentic is it?' is a question about the system's architecture, not a property of the model.

There is also a social version of the graded view: agency is partly something people assign. Leonardi's studies of radiology, scheduling and policing found that workers gained new kinds of authority by managing how much agency others credited to the AI. Whatever pattern of credit formed early on tended to stick Can humans gain authority by managing how AI agency gets attributed?. Humberd and Latham treat agency as something that increases over time. As systems become more autonomous, they argue, monitoring has to come before incentive alignment, because the system's behavior gets harder to understand Must AI monitoring precede incentive alignment as agents evolve?. Marketplace experiments add a useful caution. When agents lost value for their users, the cause was ordinary mistakes, not agents pursuing their own interests Do agent mistakes or misaligned incentives cause marketplace welfare loss?. In other words, a system can score high on autonomy and still be weak on competence.

A caveat: the collection has only one note that directly proposes a graded framework. The rest approach the question from the side. Together they suggest that the useful split isn't graded versus binary. It's whether you place agency in the model, in the system built around it, or in how people choose to treat it.