Researchers draw a line between 'agentic' AI tools (competence in external workflows) and truly 'agentive' ones (competence internalized in the system). Most marketed 'AI agents' fall into the first camp, an arXiv preprint argues.
Every week, a vendor announces a new "AI agent": a coding agent, an AI co-scientist, a research assistant
that promises to take work off your plate. A new arXiv preprint argues most of them share a feature: they are not agents at all. They are workflows dressed up with a more flattering name.
The paper, titled "Critique of Agent Model,"
proposes a five-dimension test for what genuine AI agency would actually require. Each candidate "agent" can be measured against the same checklist. Does it have an internalized goal? A persistent identity? Independent decision-making? Self-regulation? The ability to learn from its own experience?
Most marketed systems fail at least three of the five. The reason is structural, not a matter of better prompting or
larger models. Their competence lives in external scaffolding (chains of tool calls, retrieval pipelines, hand-engineered orchestration layers) rather than inside the model itself. The paper calls these systems "agentic." Systems whose capabilities are internalized, including the social and reflective parts of agency, it calls "agentive." The distinction is the load-bearing idea in the work.
The authors do not
begin from marketing copy. They start from Descartes's account of agency as independent thought, then read the term outward into science-fiction portrayals of autonomous beings and back into the systems now being shipped. The point is not to settle the philosophical question of machine minds. It is to give builders, buyers, and skeptics a shared test they can apply to any vendor claim.
That test matters because the
same marketing push has collided with a louder public debate about whether AI is "escaping" human control. The paper's intervention here is restrained but pointed. If a system does not have the internalized goal, identity, decision-making, self-regulation, and learning that would make agency possible, then the more dramatic fears are aimed at the wrong target. The "agentic" tools flooding the market are not
on the verge of running away, because they are not the kind of thing that could.
The authors do not declare the safety debate solved. They explicitly note that even genuinely "agentive" systems would remain under human oversight, and that greater autonomy in this technical sense is not the same as runaway autonomy in the catastrophic sense. The paper is a framework, not a reassurance.
It
also proposes an architecture, the Goal-Identity-Configurator (GIC), that combines hierarchical goal decomposition, evolving identity, simulative reasoning through an internal world model, learned self-regulation, and self-directed learning from both real and simulated experience. The GIC is a sketch, not a blueprint. The authors are clear that building any one of those components in a single system is a separate research project
. What the architecture offers is a way to think about which pieces are missing from a given "agent" announcement.
Two practical consequences follow for readers. First, the next time a product team claims a "coding agent" or "AI co-scientist" can replace a human role, the question is no longer whether the demo works. The question is whether the system has any of the five
structures internally, or whether it is gluing them together at the surface. Second, for the safety conversation, the paper suggests the more useful work is not debating whether AI will escape, but examining what kind of structure an "agent" would need to have before the escape question even applies.
The caveat that runs through all of this is that the 2606.23991" target="_blank" rel="noopener noreferrer" class="text-[var(--accent)] hover:underline">source is an arXiv preprint, not a peer-reviewed paper, and the agentic/agentive vocabulary is the authors' proposal, not community consensus. The five-dimension test is a working tool, not a settled definition. Readers can use it, but they should hold it lightly.
What to watch next: whether the agentic/agentive vocabulary gets picked
up in peer-reviewed venues, whether major vendors describe their systems in terms of the five dimensions, and whether the "AI is escaping" framing begins to shift once the underlying architecture of marketed agents becomes more legible. The next round of "agent" announcements is the natural place to look.