Multi-agent systems do not fail on smarts. They fail on serving, coordination and shared queues

Agents do not arrive independently

In a preprint, researchers introduce MAServeBench, a benchmark for LLM serving systems under multi-agent workloads, and start from an observation that single-agent serving benchmarks miss. Multi-agent systems generate non-independent request patterns: one agent's output triggers another's input, contexts evolve as collaboration proceeds, and requests arrive in bursts tied to the task rather than at random. Those patterns create unique concurrency, computation and memory bottlenecks in serving hardware, and, crucially, serving efficiency does not always translate into timely task completion, because a fast individual response can still wait on a coordination dependency. A serving system optimised for throughput per request can therefore look efficient while tasks finish late. m:…1065

A second preprint, INFRAMIND, treats that coupling as the central design problem. Shared serving queues create a coupled decision problem, the authors argue, in which reasoning workflows determine load and load determines which reasoning choices are affordable. Their hierarchical policy selects collaboration graphs and reasoning strategies based on real-time load and time budgets, rather than fixing a workflow in advance. In a 15-model deployment with Azure timings, the authors report that on-time correct answers rise from 11.6% to 54.8% over MasRouter, and low-load accuracy improves by up to 7.6 percentage points over the strongest evaluated baseline. The size of the on-time gain says less about any single model and more about the cost of ignoring the queue: a good plan that cannot run in time is not a good plan. m:…4212

Make skills and messages smaller, not just models

Two further preprints attack the cost of coordination itself. SkillMem replaces bulky textual or layer-wise skill representations with compact latent memories, read through a role-conditioned cross-attention reader that lets the same latent memory guide different agents according to their roles and current hidden states. The authors report storage reduced by more than 38 times for a library of 100 skills compared with alternative representations, and end-to-end inference speedups of up to 4.23 times in latent multi-agent reasoning, with behavioural distillation matching the change in token predictions induced by a textual skill while keeping the backbone model frozen. The practical point is that a skill library is infrastructure: if every skill is a long prompt, every agent pays for it on every turn. m:…6612

Similarly, a preprint on latent multi-agent communication proposes bypassing natural language exchange between agents altogether. Its SelectKV method reduces communicated key-value positions and logical payload by 9.7 to 9.9% for standard Transformers, while HyLaMAS, for hybrid linear-attention models, reduces decoded output tokens by 53 to 80% and achieves speedups of 1.16 to 2.78 times relative to text-based multi-agent communication. Agents talking to each other in full sentences is convenient for debugging and expensive for serving, and these papers suggest the industry will increasingly treat inter-agent language as a cost to be optimised, not as a given. m:…8985

Buy the orchestration, not just the models

For an executive deciding where to invest in an agent deployment, these preprints point to a sequencing decision. Better individual models will not fix a serving layer that treats agent requests as independent, a workflow that ignores queue load, a skill library stored as long text, or agents that coordinate by writing essays to each other. The gains reported here, from 11.6% to 54.8% on-time correct answers, 38-fold storage reduction, and up to 80% fewer decoded tokens, all come from the layer around the models. That layer is less visible in a demo than a new model, and it is where a multi-agent deployment most often succeeds or fails in production. m:…4212 m:…6612 m:…8985