What question did this study set out to answer?

The study aims to define Preservation-Bounded Autonomy (PBA) to ensure AI agents maintain cognitive preservation while evolving.

March 3, 2026Open Access

Preservation‑Bounded Autonomy (PBA): Cognitive Preservation as a Boundary Condition on Self‑Evolving AI Agents

Puntos clave

The study aims to define Preservation-Bounded Autonomy (PBA) to ensure AI agents maintain cognitive preservation while evolving.
Introduced a formal definition of Preservation-Bounded Autonomy (PBA) as a governance paradigm for AI agents.
Developed a two-dimensional autonomy design space to categorize agent autonomy.
Reframed existing AI safety laws to include cognitive preservation as a boundary condition.
Outlined measurement dimensions for cognitive preservation and proposed a research agenda.
Positioned PBA against existing frameworks and defined DAEDALUS as a reference implementation.
PBA model shows high autonomy for agents while dynamically constraining their operational autonomy to preserve human cognitive capabilities.
Identified three autonomy zones revealing a gap in current systems not occupying the PBA zone.
Proposed four dimensions for measuring cognitive preservation, indicating the need for further research and validation.

Resumen

Extended abstract introducing Preservation-Bounded Autonomy (PBA) a design paradigm for AI agent autonomy governance where an agent's operational autonomy is dynamically modulated by non-evolvable structural invariants that preserve human independent task capability. PROBLEM:Self-evolving AI agents autonomously optimize models, memories, tools, and workflows. Existing safety frameworks (e.g., Three Laws of Self- Evolving Agents: Endure, Excel, Evolve) specify constraints in terms of AI-internal properties. They do not address whether autonomous optimization degrades the human partner's independent capability, a failure mode we term "cognitive atrophy by delegation creep." Empirical evidence shows behavioral safety constraints are brittle under self-evolution. CONTRIBUTION:(1) Formal definition of Preservation-Bounded Autonomy (PBA): an agent possesses high-autonomy capabilities but its operational autonomy is dynamically constrained by non-evolvable structural invariants enforcing cognitive preservation. (2) A two-dimensional autonomy design space (Operational Autonomy A × Cognitive Preservation Index P) that separates capability from governance, exposing three regions: Reactive Zone, Unconstrained Autonomy Zone, and PBA Zone. No surveyed system occupies the PBA zone. (3) Reframing of the Engage Law (Leu, 2026) as a boundary condition on the "Evolve" objective in self-evolving agent frameworks, identifying unbounded self-evolution in human-coupled systems as an incomplete safety specification. (4) Framework positioning against SOLA (Liu et al., 2023), Generative Agents (Park et al., 2023), Autonomy Levels (Feng et al., 2025), and MASE/Three Laws (Fang et al., 2025). (5) Operationalization agenda: four measurement dimensions for cognitive preservation (verification reconstruction rate, autonomous-to-confirmedaction ratio, override trajectory, reasoning independence test), defined as research program, not validated instruments. (6) DAEDALUS as reference implementation, with explicit connection to the AXIOM architecture (structural invariants for physical validity → cognitive validity), establishing a coherent research program. RELATED DEPOSITS:- Engage Law position paper: doi:10.5281/zenodo.18661084- ARIADNE working memory specification: doi:10.5281/zenodo.18506520- Convergence matrices: doi:10.5281/zenodo.18507521- Dual-memory architecture: doi:10.5281/zenodo.18507663 Parent project: DAEDALUS (Cognitive AI Agent Architecture)Institution: University of Oradea, Romania

Leer artículo completoexternamente

Me gusta

Guardar

Ver artículo completo