Agentic AI systems differ in a key way: who or what actually holds the 'agency' — the model, the tools, the system, or the humans overseeing it?

This explores how different ways of classifying agentic AI systems decide where the 'agency' actually sits: in the model, the tools, the surrounding infrastructure, or the humans and protocols around them. One caveat first: the collection has no single paper laying out 'four archetypes' of agentic systems. What it does have are several competing four-part (and three-part) breakdowns. Each one puts agency in a different place, and comparing them is more revealing than any one of them alone.

The first framing puts agency on a dial instead of in a box. OpenAI's framework scores systems on four independent axes: how complex the goals are, how complex the environment is, how well the system adapts, and how much it executes on its own Should we define AI agents as a spectrum rather than a category?. Under this view, agency isn't located anywhere specific. It is a profile, so a system can be highly autonomous yet barely adaptive. A second four-part scheme is a 2x2 grid for how agentic systems learn and improve. One dimension asks what gets optimized (the agent itself or the tools it calls). The other asks what the feedback signal is (whether execution succeeded or whether the final answer was good) How do agentic AI systems decompose into adaptation paradigms?. Here agency is split between the agent and its tools, and the scheme makes you choose which one you're improving. That choice explains practical trade-offs, such as a better search query versus a better final answer.

A third view drops the four-way split and stacks agency in layers: the model's trained reasoning, the harness (the system code that turns model outputs into actions), and the code the agent writes for itself while it works What are the three distinct layers of agent code?. Each layer fails differently, so 'where is the agency?' becomes the practical question 'where do I intervene when it breaks?' This connects to the argument that code itself works as the agent's operating medium, because it can be run, inspected, and carries state Can code serve as the operational substrate for agent reasoning?. In that picture, part of the agency lives in the artifacts the agent produces.

The most surprising material is about where people *perceive* agency to be, and how that can be redesigned. When an AI system is shown to people as a language model plus a tool-using agent layer, they assign more blame to the agent layer simply because it sits closer to the outcome Does decomposing AI systems shift where people assign blame?. Agency can also flow back toward humans: agentic systems can assign tasks *to* people, which reverses the usual direction of delegation Can AI systems delegate tasks back to humans?. Executives are split along similar lines, with 76% treating agentic AI as a coworker even while managing it as an asset Does agentic AI fit the tool-or-coworker framework?.

For multi-agent systems, the lesson is that agency located on paper is not agency in practice. Labeling one agent as 'accountable' changes nothing unless the workflow actually routes the work through that agent Does accountability placement actually shape multi-agent outcomes?. The best-performing setups mix the two approaches: a fixed external order of turns, with agents choosing their own roles inside it. That beat fully centralized teams by 14% and fully self-organizing ones by 44% Do self-organizing agent teams outperform rigid hierarchies?. Across all of these, the takeaway you might not expect is that you can't settle where agency lives just by drawing a diagram. It shows up in how control actually flows through the system, and in where observers decide to look.