Efficiency in LLM Multi-Agent Systems: A Survey

Large language model (LLM)-based multi-agent systems (MAS) have moved beyond small debate ensembles toward workflows in which specialized agents operate across tools, memory, and external environments. This shift expands what a system can search and verify, but it also makes collaboration expensive: repeated inference and stateful coordination can outweigh the performance gain. The central efficiency question is therefore not simply whether multiple agents outperform a single agent under a fixed budget, but when the benefit of collaboration justifies its end-to-end resource cost. Exist-ing work addresses this question from different aspects: some studies decide whether collaboration should be invoked, while others assume a multi-agent workflow is already in place and seek to reduce its overhead. Since these studies define baselines and costs differently, their reported efficiency gains are difficult to compare. This motivates a unified accounting view that relates each performance gain to the collaboration overhead required to obtain it. This survey organizes these efforts around a common view of the problem. We first characterize the Collaboration Boundary, which reviews previous works and determines when collaboration is worth invoking. We then organize existing efficiency methods by three primary control points: Topology, which determines the participating agents and communication structure; Runtime, which controls the information, models, and state activated during execution; and Optimization, which learns reusable prompts, workflows, and policies through search or training. In Evaluation, we further examine how these methods should be evaluated through matched baselines, explicit resource accounting, and lifecycle cost. We further provide a Synthesis that compares the resulting evidence, clarifies where reported savings are well supported, and explains why some efficiency gains may not transfer across settings. We conclude with Open Problems that identify the remaining gaps in- efficient and scalable MAS design. Across the literature, a consistent picture emerges. MAS efficiency is a quality–resource trade-off, not a single measure of agent count, tokens, or cost. The best design depends on the task, the system bottleneck, and the stage at which resources are spent.