The Distinction Between Agentic and Agentive Systems

The authors argue that the current industry trend of labeling LLM-based tools as "agents" is misleading. Most existing systems are agentic, meaning their competence is derived from external scaffolding, hard-coded workflows, and prompt-engineered chains. These systems are designed for specific, prescribed tasks and lack internal autonomy.

In contrast, agentive systems possess endogenous capabilities. Their ability to reason, interact, and adapt arises from within the model architecture itself. The authors posit that genuine agency is defined by five internalized dimensions: goal, identity, decision-making, self-regulation, and learning. Without these, a system is simply a sophisticated automation tool, not an autonomous agent.

The GIC Architecture for General-Purpose Agents

To move beyond current limitations, the authors propose the Goal-Identity-Configurator (GIC) architecture. This framework aims to shift AI from task-specific automation to open-world autonomy through:

• Hierarchical Goal Decomposition: The ability to break down high-level objectives into actionable steps internally.
• Identity Evolution: Maintaining a consistent "self" that evolves based on experience, rather than resetting per session.
• Simulative Reasoning: Grounding decision-making in a separately trained world model that allows the agent to simulate outcomes before taking action.
• Learned Self-Regulation: Moving away from hard-coded guardrails toward a system that learns to regulate its own behavior.
• Self-Directed Learning: Enabling the agent to learn from both real-world feedback and its own internal simulations.

Implications for Safety and Control

The paper emphasizes that as systems move toward true agency, the traditional methods of human oversight (like simple prompt-based constraints) will become insufficient. The authors suggest that auditability and controllability must be built into the agentive architecture itself. By ensuring that goals and identities are transparent and modifiable, developers can maintain human oversight even as systems achieve higher levels of autonomous decision-making in open environments.