This paper proposes a compact way to think about large groups of interacting agents — people, robots, neurons, or software programs — when their actions change the information they and others see. The author builds the framework from two agent-level ingredients. “Power” means how much influence one agent has on collective outcomes. A “response function” describes how an agent changes its actions when it sees a collective signal. From these two ingredients the paper shows how big-picture properties can emerge.

Using those agent-level pieces, the paper derives system-level quantities such as total power, useful power (the part of power that produces desired outcomes), entropy (a measure of disorder), order, fragility (tendency to fail), and mobility (ability to change). The work also defines a system utility function that includes a risk-appetite coefficient — a single parameter that captures whether the system prefers growth or safety. With that utility the paper finds an optimal degree of order that trades off productivity, stability, and adaptability.

At a high level the framework focuses on feedback loops: agents act, their actions change what everyone observes, and those observations change future actions. A clear conclusion is that stronger synchronization — when many agents act the same way — can boost collective output. But synchronization can also concentrate influence and make the whole system more fragile and less able to move in new directions. The paper also argues that ideas like order, entropy, and useful energy depend on the task and on whose goals you consider — they are not absolute.

The paper places this framework alongside existing tools in economics, physics, and computer science. It relates to mean-field game ideas and stochastic games (which simplify many agents by averaging their effects), to dynamical systems (equations for how states change over time), to agent-based modeling (simulating many different agents), and to control and learning theory. The author suggests applications ranging from financial markets and social networks to brain networks and multi-agent artificial intelligence, including large language model (LLM) based systems.