Defining True AI Agency Beyond Engineered Workflows

The proliferation of AI systems marketed as "coding agents" or "AI co-scientists" has blurred the lines between automation and genuine agency, raising questions about both system capabilities and potential existential risks. This paper critically examines the current landscape of AI agents, drawing on philosophical concepts and science fiction portrayals to distinguish between engineered workflows and true autonomy. The authors argue that genuine agency necessitates the internalization of structures related to goals, identity, decision-making, self-regulation, and learning within the system itself, rather than relying on external scaffolding. They differentiate between "agentic" systems, whose competence stems from predefined workflows, and "agentive" systems, whose capabilities, including social interaction, emerge endogenously. This distinction defines the boundary between systems designed for specific tasks and those capable of operating autonomously in the open world. Building on this analysis, the paper proposes the Goal-Identity-Configurator (GIC) architecture for a general-purpose agent model. This architecture combines hierarchical goal decomposition, evolving identity, simulative reasoning grounded in a trained world model, learned self-regulation, and self-directed learning from both real and simulated experiences. The work also offers insights into the auditability, controllability, and safety of these more autonomous "agentive" systems, emphasizing the importance of human oversight.