The Autopoietic Agent is an early research project exploring whether artificial general intelligence should be designed less like an unbounded optimizer and more like a living system that must continuously maintain its own operational stability.
The working paper draws from biological processes such as neuroplasticity, the hypothalamic-pituitary-adrenal stress axis, starvation metabolism, nociception, and homeostasis. It interprets these systems as computational strategies for learning efficiently, allocating limited resources, responding to threats, and remaining resilient under changing conditions.
The proposed Autopoietic Architecture centers on an Interoceptive Network that models and regulates the agent’s internal state. Instead of pursuing a single, unbounded utility function, the agent operates within homeostatic limits and adapts its behaviour to restore dynamic equilibrium. The paper introduces early concepts including Systemic Perturbation Indexes, Efficient Inference Modes, Dynamic Network Sparsification, and layered reflexive defence mechanisms.
This is a theoretical blueprint, not a validated AGI system. The next meaningful stage is empirical: formalising the architecture, implementing simulations, and testing whether its proposed mechanisms improve resource efficiency, resilience, corrigibility, and alignment under measurable conditions. Some claims may need to be revised or discarded as evidence emerges.
For now, the paper is a research direction and an invitation to investigate a central question: can bounded self-regulation and embodied vulnerability provide a stronger foundation for autonomous intelligence than endless optimisation?
Links: