source: arxiv artificial intelligence: critique of agent model
level: research
researchers from multiple institutions have published a critique of current ai agent models, drawing on descartes' idea of independent thought and science fiction portrayals of autonomy. they argue that many systems marketed as agents, like coding assistants or ai co-scientists, lack true agency because their core functions remain externally driven. the paper proposes five dimensions to evaluate agency: goal, identity, decision-making, self-regulation, and learning.
the authors contend that for a system to be genuinely agentic, these five elements must be internalized within the system itself. for example, a true agent would set its own goals rather than simply following human instructions, maintain a persistent sense of identity, make decisions autonomously, regulate its own behavior, and learn from experience without external reprogramming. current large language model systems typically fall short because their goals and decision-making frameworks are imposed by human designers.
this framework aims to clarify the boundary between automation and agency, which has practical implications for building more capable systems and for assessing risks. the paper suggests that fears about ai escaping human control are tied to the degree of internalized agency a system possesses. by defining agency more precisely, the researchers hope to guide development toward safer and more effective ai tools while avoiding unnecessary alarm about systems that are merely advanced automation.
why it matters: understanding true agency helps ai developers build safer systems and allows data scientists to better evaluate the capabilities and risks of agentic tools.
source: arxiv artificial intelligence: critique of agent model