Cybernetics as Theoretical Foundation for LLM-Based Agents

A new paper on arXiv (2605.10754) argues that cybernetics, the mid-20th-century science of control and communication, provides the missing theoretical framework for LLM-based foundation agents. These agents, which perceive, reason, and act across thousands of reasoning steps, are becoming dominant for open-ended, long-horizon tasks. The field is currently engineering-driven, with primitives like tool loops, memory banks, and reflection steps assembled by trial and error. The authors pose fundamental questions about on-task behavior, representational capacity limits, and safe self-improvement. They map six canonical cybernetic principles to foundation agents, offering a theoretical scaffold for the field.