Add a model to your AI council and you'd assume the cost goes up by one model. It doesn't. It goes up by every conversation that model now has with the others. That's the part the bill hides.
Most multi-agent debate systems are fully connected — every agent broadcasts to every other agent, every round. So the cost isn't the headcount. It's the handshakes. And handshakes grow as the square of the headcount.
The number that surprises people
For a system of n agents running R rounds of fully connected debate, the message count is O(Rn²). Six agents over three rounds isn't six units of work. It's ninety messages, each carrying a chunk of reasoning that turns straight into tokens and latency. Double the agents to twelve and you don't double the cost — you roughly quadruple the message traffic. That's why "just add another model" stops being free fast, and why large councils that look great in a demo get quietly retired in production.
A new arXiv paper out June 1 — Dynamic Trust-Aware Sparse Communication Topology for LLM-Based Multi-Agent Consensus, arXiv:2606.01828, which calls its method DySCo (Dynamic Sparse Consensus) — names this directly. Fully connected communication makes messages, token cost, and latency all grow approximately quadratically with the number of agents. The paper's fix is not what most people reach for.
The fix isn't fewer models. It's fewer conversations.
The instinct, when a council gets expensive, is to drop models. DySCo keeps the models and drops the broadcasting. Each round, instead of letting all six agents talk to all five others, it picks a small number of high-value edges — at most k neighbors per agent — and routes messages only along those. The expensive thing was never the council. It was making everyone talk to everyone.
The trick is in how it decides which conversations are worth having. DySCo scores each candidate edge on three signals: how reliable an agent has been so far, how much its answer diverges from the others, and how relevant it is to the task at hand. Then it spends its message budget on the edges that score highest. An agent confidently agreeing with the room adds nothing — that edge gets pruned. An agent holding a different answer for a defensible reason is exactly the conversation you want to pay for — that edge stays.
It also compresses what gets said. Rather than broadcasting a full reasoning trace, each message carries three things: the agent's current answer, the single most critical reason supporting it, and one counterexample or revision suggestion for the receiver. Then it aggregates answers by trust weight and stops early once the council stabilizes, instead of grinding through every scheduled round.
Pruning blindly versus pruning by value
Cutting communication to save money is not new. You can wire agents into a fixed ring, or have each one talk to k random others. Both cut cost. The paper tests them — Static Ring, Random-k — and they do reduce overhead. They also leave accuracy on the table, because they prune without looking. A fixed topology keeps low-value chatter while dropping the one cross-check that would have caught the error.
That's the distinction worth holding onto. Static sparse topologies cut the conversations cheaply. DySCo cuts the conversations that don't matter and keeps the ones that do. Same budget, very different result.
The measured gap, on gpt-3.5-turbo across GSM8K, LogiQA, and StrategyQA (300 instances each, six agents, three rounds, two neighbors): compared with fully connected debate, DySCo cuts token consumption by roughly 70% and latency by nearly half — and it's more accurate, not less, reaching 84.3% on GSM8K. The complexity drops from O(Rn²) to O(Rnk). Cheaper, faster, and better at the same time only happens when the thing you removed was waste.
Why this is a council problem, not a footnote
There's a version of this finding that reads as a narrow efficiency tweak. The more useful reading: it changes what a council can be. As long as adding a model means adding a full round of conversations with every other model, councils are capped at small sizes by economics alone. Make the conversations adaptive and the cap moves. A council can carry more perspectives without the bill scaling like a square.
DySCo isn't alone in pointing here. A spectral-diagnostic paper from earlier this spring (arXiv:2605.11453) gives a way to predict, before you run anything, which communication structures will fail. The Aegean work (arXiv:2512.20184) gives a principled rule for when a council can stop early. DySCo sits between them: a per-round, per-task rule for which edges are worth the tokens. Together they're a toolkit for treating a council's wiring as something you design, not something you accept.
That's the part that maps onto how we think at Shingikai. A Traditional Council, a Round Robin, a Red Team vs. Blue Team run aren't just different prompts — they're different answers to the question DySCo is asking: which agents should be in conversation, and when. The strategy you pick is a communication topology, chosen for the question in front of you.
One model gives you an answer. A council gives you a position — but only if the agents that disagree actually reach the ones who can be moved. The cost of a council was never the number of voices in the room. It was how many of them you forced to talk over each other.
Try it free — no signup. shingik.ai