A multi-agent system does not need every agent: picking the right ones matters more

Broadcasting a task to every agent looks safe: what if one of them happens to have the skill you need. But along with the number of participants come rising token costs, latency, duplicated answers and the likelihood that one agent's error will spread further.

The authors propose assembling a coalition for the specific task. The system weighs not only the benefit of a new agent but the price of activating it, and then decides who else needs to be brought into contact. That is, what gets optimised is not an abstract number of agents but the team's net value: the result minus the cost of participants and communications.

Shapley estimates help with that: they approximately show how much a particular agent can raise the overall result. In synthetic experiments the greedy router obtained 99.5% of the optimal utility from exhaustive search while activating an average of 1.96 agents out of 8. For comparison, the full-broadcast mode reached only 38.8%.

The practical conclusion here is fairly down to earth: in a multi-agent system it is worth optimising routing first and only then adding new specialists. Sometimes quality grows not from widening the team but from not giving the surplus agents the floor at all.

But this is no ready universal guarantee of savings. The main results come from synthetic data; under a strong violation of submodularity or with noisy quality estimates the figure fell to 66%. The proven guarantees apply only to special variants of the problem, and the main router remains a heuristic. The authors also propose testing the approach separately on real benchmarks of multi-agent LLM systems — so the question of transferability is still open.

If you are building a system of several agents, what costs more right now: bringing in another specialist, or giving the agents you have already chosen too much communication? Source: cs.AI updates on arXiv.org