At what network size do frontier models stop coordinating?
At what network size does adding an agent stop helping?
Opening
Ask whether a multi-agent system works and most teams point at a reasoning score: the system solved the task, therefore the system is fine. But a score tells you the answer arrived, not how it got there — whether the agents used the network they were given, agreed on anything, or talked past each other. That distinction is invisible on standard benchmarks, and it is the one that decides whether a fleet of agents is an asset or an expensive queue. AgentsNet was built to measure coordination itself, at sizes no earlier benchmark attempted.
What the paper does
AgentsNet draws inspiration from classical problems in distributed systems and graph theory, and measures the ability of multi-agent systems to collaboratively form strategies for problem-solving, self-organization, and effective communication given a network topology. The evaluation covers a variety of baseline methods, including homogeneous networks of agents which first have to agree on basic protocols for organization and communication before any task work begins. Where existing multi-agent benchmarks cover at most 2-5 agents, AgentsNet is practically unlimited in size, so the authors probe frontier models in a setup with up to 100 agents and watch what happens as the network grows.
Findings
Standard scores cannot see coordination — Reasoning benchmarks report whether the task was solved and stay silent on whether the topology was used at all. The gap is not measurement noise; it is a quantity existing benchmarks simply do not report.
it is unclear whether these systems are able to leverage their topology effectively
What the benchmark measures — Strategy formation, self-organization, and communication are scored inside a given network topology, so a system can no longer win on reasoning alone while its coordination goes unmeasured. The topology becomes part of the test.
measures the ability of multi-agent systems to collaboratively form strategies for problem-solving, self-organization, and effective communication given a network topology
Small networks work; scale breaks them — Frontier LLMs already demonstrate strong performance on small agent networks, then begin to fall off once the size of the network scales. What degrades is coordination, not raw capability, and the benchmark was built to keep watching that curve as models improve.
strong performance for small networks but begin to fall off once the size of the network scales
Caveat
The stated limitation is the measurement gap AgentsNet was created to fill: standard benchmarks indicate how well multi-agent systems solve reasoning tasks, but it is unclear whether these systems are able to leverage their topology effectively. Until a benchmark reports that quantity, a good score cannot be read as good coordination.
What this means for 5ac
This is the measurement behind the question every customer asks: does adding another profile help? Before scaling a pipeline, the board should record not just whether the card completed but whether the agents actually used the wiring — who read whose comment, which parent the child consulted. The fall-off with size is the operating rule: keep the working network small, and add a role only when a run shows coordination failing for lack of one, not for lack of headcount. When a card genuinely needs more hands, split it into two small networks instead of growing one, because on this evidence the gains leave as the network grows.
Source
- AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs — https://arxiv.org/abs/2507.08616