How do coordination architectures trade off efficiency against resilience?

Within the same 1994 survey, Malone and Crowston use a "generic task assignment problem" to make the dependency-management framework concrete. Citing formal models from Malone [1987] and Malone and Smith [1988], built with "techniques from queuing theory and probability theory," they report: "the centralized schemes had lower coordination costs, but were more vulnerable to processor failures. Decentralized markets, on the other hand, were much less vulnerable to processor failures but had high coordination costs. And decentralized hierarchies ('product hierarchies') had low coordination costs, but they had unused processor capacity which led to high production costs."

The task assignment problem behind the comparison: "clients" discover tasks only as work proceeds, and only specialized "server" actors can perform each task type. The centralized mechanism routes all information to "a central decision maker who decides which servers will perform which tasks." The decentralized alternative, from "the competitive-bidding scheme for computer networks formalized by Smith and Davis [1981]," has a client broadcast a description of the activity and required qualifications, servers submit bids describing "their qualifications and their availability," and the client award the task directly — a market with no central assigner. What separates the three architectures in the models is this three-way cost split (production, coordination, vulnerability), not any difference in the individual servers' task competence.

This is a decades-older, formally modeled version of the centralized-versus-decentralized tradeoff that Do teams of personal agents outperform a single coordinator? finds empirically in frontier-model agent teams: a single coordinator beating uncoordinated per-user agent teams is an agent-era instance of "centralized schemes had lower coordination costs." It also supplies a formal-modeling precedent underneath Does AI adoption require rethinking how organizations coordinate?, whose modularity argument restates the same tradeoff in 2021 terms — a modular (decentralized) organization trades away synergy to cut coordination cost, where this paper's models trade coordination cost against vulnerability and production cost instead. When does adding more agents actually help systems? names a comparable tradeoff for present-day agent topologies, where this paper supplies the original formal vocabulary — production, coordination, and vulnerability costs — for weighing architecture choices against each other.

These are queuing-theoretic formal models of abstract clients and servers moving discrete tasks, summarized secondhand inside a 1994 survey — not a study of software agents, LLMs, or any deployed organization, and "production costs," "coordination costs," and "vulnerability costs" are modeled constructs, not measured dollar figures. The paper does not claim any one architecture dominates: citing Gurbaxani and Whang [1991], it notes "IT can lead to either centralization or decentralization, depending on how it is used." The implication for agent system design is a tradeoff to weigh deliberately against the dependencies a given task actually has, not a universal argument for one coordination topology over another.