When people tune a multi-agent setup, they almost always reach for the same knob first: how much should each agent remember. Give the models more context, the thinking goes, and they'll coordinate better. A new arXiv paper says that knob has no fixed sign.
Aliakbar Mehdizadeh and Martin Hilbert at UC Davis ran 432 simulations of a networked Naming Game — a classic coordination task where agents have to converge on a shared label — across eight fixed 16-agent topologies, varying two things: how deep each agent's memory ran, and how the agents were wired together. The headline result is the kind of thing that should stop anyone building a council: longer memory slows the time to settle in decentralized networks but accelerates it in centralized ones. Same parameter, opposite directions, depending entirely on the shape of the network it's running in.
That alone kills the idea that memory is a free dial you turn up until things get better.
"Settling" is not the same as "agreeing"
Here's the part that matters most for anyone who thinks a tighter, more centralized council is the safer bet. In the centralized topologies, "faster settling" doesn't mean the system reached one answer. It means the system locked into a fragmented plateau — competing conventions frozen in place — more quickly. The network stops moving. It just doesn't stop on consensus.
A centralized council settles faster. It just doesn't settle on one answer.
The paper documents this as a memory-mediated speed-unity trade-off: centralized networks consistently preserve more competing conventions than decentralized ones, but how fast they get there depends sharply on memory. So the two design choices everyone treats as independent — wiring and memory — aren't independent at all. The authors are blunt about the implication: memory depth and communication topology should be co-designed, not optimized in isolation.
If you've ever tuned one of those settings without thinking about the other, this paper is telling you that you were turning a knob whose effect you couldn't predict.
The bridge agent takes a penalty
The second finding is the one that should make anyone designing a Chairman-shaped council uncomfortable. Inside each network, the authors looked at which agents coordinated well and which didn't. The agents sitting on high-betweenness bridges — the connector nodes, the ones routing information between clusters — suffered what the paper calls a brokerage penalty. They coordinated worse. Meanwhile agents tucked inside locally clustered neighborhoods achieved higher coordination success.
That cuts against the intuition behind a lot of multi-agent design. The instinct is to route everything through a central hub: one synthesizer that sees all the perspectives and resolves them. This result says the hub position is not automatically the strong one. Being the broker between groups can make coordination harder, not easier — the agent in the middle is exactly the one absorbing the most conflicting signal.
This doesn't mean a synthesis layer is wrong. It means the synthesizer's job is harder than the org chart suggests, and the structure around it determines whether it converges the room or just freezes the disagreement in amber.
Why a council builder should care which way this cuts
Map the topologies onto how councils actually run and the lesson gets concrete. A centralized, Chairman-as-star structure is one strategy. A decentralized, everyone-talks-to-everyone structure — a round-robin or a mesh — is another. This paper is empirical evidence that the choice between them materially changes the outcome, and not in the direction you'd guess. Centralized structures preserve divergence; decentralized ones drive harder toward a single convention.
So the right question isn't "which topology coordinates best." It's "do I want this council to force one answer, or to surface and hold the disagreement?" Those are different jobs, and they call for different wiring.
For a value-laden question — an ethics call, a strategy bet, a decision where reasonable models should disagree — a structure that preserves competing conventions is a feature, not a bug. You want to see the two camps, not get them blended into a confident average that hides the split. For a factual question with one right answer, you want the structure that converges. The same machinery, pointed at the wrong job, gives you the wrong shape of output.
This is the whole argument for having more than one council strategy instead of one fixed pipeline. Chairperson Synthesis and Red Team vs Blue Team exist precisely to preserve and route disagreement; Round Robin and a tighter council exist to drive toward convergence. The Mehdizadeh-Hilbert result is the structural reason those aren't interchangeable.
One more thing worth noting
In looking for a tractable mechanism behind all this, the authors found that the agents' choices were well captured by Fictitious Play — meaning the agents adapt based on what they believe other agents will do, not on rewards they've collected. Belief-based, not reward-based. That's a small line in the paper with a large consequence: it says these systems coordinate the way game theorists have modeled human convention-formation for decades, which makes the whole body of social-coordination math fair game for predicting how a council behaves before you run it.
The takeaway isn't "centralize" or "decentralize." It's that the structure of a council and the memory of its members are a single design surface, and the output you get — one answer or a preserved split, fast or slow — falls out of how you shape that surface. Pick the shape for the decision, not by habit.
If you want to watch what a council does with a question where the models genuinely disagree — and decide for yourself whether you want that disagreement resolved or shown — try it free, no signup. shingik.ai