Critique of Agent Model

Modern AI systems are frequently labeled as agents, yet they often function as sophisticated software pipelines. The authors argue that current systems are merely agentic—relying on external prompts, predefined tools, and rigid workflows to complete tasks. In contrast, they define agentive systems as those that possess genuine autonomy, where capabilities arise endogenously from the model itself rather than from external orchestration. This distinction is crucial for both the development of capable AI and for addressing safety concerns regarding machine autonomy.

To move from agentic to agentive systems, the authors analyze agency across five fundamental dimensions: goals, identity, decision-making, self-regulation, and learning. They argue that for a system to be truly agentive, these structures must be internalized. For example, rather than following a fixed, externally provided task list, an agentive system should be capable of hierarchical goal decomposition, where it breaks down long-term objectives into actionable subgoals. Similarly, identity should be an evolving latent variable that allows the agent to update its self-model based on experience, rather than remaining static.

The authors introduce the Goal-Identity-Configurator (GIC) architecture as a blueprint for general-purpose agent models. This architecture separates the agent model (which decides what to do) from the world model (which predicts what will happen). The GIC integrates several key components:

• Hierarchical goal decomposition for long-term planning.
• An evolving identity that adapts without requiring retraining.
• Simulative reasoning that uses an internal world model to evaluate future possibilities.
• A configurator for self-regulation, determining when to deliberate versus when to act reactively.
• Self-directed learning from both real-world and simulated experiences.

By keeping the agent model and world model functionally distinct, the authors aim to improve the reliability of planning and the auditability of the agent's decision-making process.