Safe AI should be bounded and multi-agent

Preprint

Major developments in frontier AI systems over the last decade have been driven by the scaling paradigm, which treats resource constraints as key obstacles to be overcome in the pursuit of more capable systems. Here, we argue for a complementary paradigm that embraces these constraints — together with the multi-agent, distributed nature of real-world deployments — as a route towards safe and scalable AI. Rather than scaling individual agents alone, we posit that legibly composing agents while deliberately bounding their capabilities, affordances, and resource budgets can reliably yield system-level competence. We call such systems bounded multi-agent systems (BMAS). Our position is that bounded agency should be a foundational principle for scaling towards safe, robust, and equitable AI. This motivates a research agenda to formally characterise bounded agency, design legible interfaces and institutions for agent ecosystems, and evaluate when bounded modular systems are more appropriate than monolithic systems.

Authors: Hyland, D; Ornia, DJ; Bishop, N; Dyer, J; Macmillan-Scott, O

• Publication status: Not published
• Peer review status: Not peer reviewed
• Language: English