Researchers have published a framework that decides, before running a full multi-agent debate, whether a task is hard enough to deserve one.
Multi-Agent Debate systems work by running several AI agents through iterative rounds of reasoning and criticism until they converge on an answer. Current frameworks apply the same expensive process to every task regardless of difficulty. HCP-MAD replaces that with a three-tier escalation system: two agents built on different base models first check for quick consensus, and if they agree, the debate ends there. If not, those same two agents enter an adaptive critique loop. Only tasks that survive both stages get handed to a larger voting pool. Across six benchmarks, the approach improved accuracy while cutting token usage compared to standard debate methods.
Most tasks are easy. Running a full committee on routine queries is waste, and that waste is a real barrier to deploying multi-agent reasoning at production scale. The choice to use agents from different base models for the initial check is deliberate — if both agents come from the same model, early agreement might just mean they share the same blind spots, not that the answer is correct. Heterogeneous disagreement is a signal; heterogeneous agreement is worth trusting.
Whether those benchmark gains hold in messier real-world deployments is the open question. Benchmarks reward tasks with clean correct answers; triage systems depend on consensus being meaningful. Those two things do not always line up outside a test set.