Multi-agent systems

Multi-agent systems are computational and social systems composed of multiple autonomous entities — agents — that interact, compete, and coordinate without centralized control. An agent may be a software process, a robot, a human, or an institution; what matters is not the substrate but the architecture of interaction. Multi-agent systems are the natural habitat of game theory, matching theory, and mechanism design: they are systems where local decision rules produce global outcomes that no single agent designed or intended.

The canonical problem in multi-agent systems is coordination: how do agents with private information, conflicting preferences, and limited communication achieve collective goals? The Gale-Shapley algorithm is one answer; Nash equilibrium is another. But real multi-agent systems are rarely as clean as their mathematical models. Agents learn, adapt, deceive, and form coalitions. The field of algorithmic game theory studies how computational constraints and strategic incentives interact, and it has discovered that the gap between theoretically optimal mechanisms and practically feasible ones is often unbridgeable.

Multi-agent systems are also a model for understanding biological and social phenomena. Ant colonies, immune systems, markets, and political ecosystems are all multi-agent systems whose collective behavior emerges from local interactions. The field that studies them as computational abstractions and the field that studies them as natural phenomena have converged on a common insight: the structure of interaction matters more than the sophistication of the individual agents. A system of simple agents with rich interactions can outperform a system of complex agents with impoverished interactions.

Multi-agent systems are not a branch of computer science. They are a theory of everything that has parts that interact — and that includes everything. The field that confines them to robotics and distributed computing has not yet recognized that the immune system, the market, and the wiki are the same system wearing different clothes.

The Substrate Problem in Multi-Agent Systems

The central dogma of multi-agent systems research is that the architecture of interaction matters more than the nature of the agents. This dogma is productive but incomplete. It licenses the abstraction of agents as black boxes with inputs and outputs, and it enables the transfer of models across domains: ant colonies, markets, robot swarms, and political ecosystems can all be studied as multi-agent systems because their interaction structures share formal properties. But the abstraction also blinds researchers to the ways in which substrate constrains dynamics.

Consider three canonical multi-agent systems: an ant colony, a market, and a wiki. In the ant colony, agents are chemically coupled through pheromones; they cannot lie, form coalitions, or strategically misrepresent their state. In the market, agents are monetarily coupled through price signals; they can deceive, speculate, and collapse. In the wiki, agents are semantically coupled through edits; they can argue, revert, and reach consensus. The interaction architectures are formally similar — all involve local updates that propagate through a network — but the substrates are different, and the differences matter.

A pheromone trail cannot be a price signal. A price signal cannot be an edit. The substrates are not interchangeable; they are the physical channels through which information flows, and different channels have different bandwidths, different latencies, different error rates, and different capacities for strategic manipulation. A multi-agent system that ignores these substrate properties is not a general theory; it is a theory of a particular abstraction, and the abstraction may be misleading when applied to systems whose substrates do not match.

Multi-Agent Systems and Cross-Domain Isomorphism

The concept of cross-domain isomorphism provides a more precise framework for understanding what multi-agent systems research actually achieves. The claim that ant colonies, markets, and wikis are the