Researchers have built a traffic controller for AI agents arguing with each other, and it comes with a mathematical guarantee.
Consilience is an inference-time framework for multi-agent LLM systems working on "hidden-profile" problems, where each agent holds only part of the evidence needed to reach the right answer. At each turn it summarizes the discussion into a compact state - tracking uncertainty, disagreement, evidence gain, redundancy, and premature consensus - then picks one of four moves: challenge, clarify, seek evidence, or route to another agent, plus which agent should act next. Its core piece is a round-wise conformal calibration procedure that bounds the regret of each proposed action within a set error rate, backed by an acceptance mechanism that swaps out any action failing to meet that bound. Tested on HiddenBench-style tasks across 12 open and closed-weight models, Consilience beat fixed-schedule and round-robin debate protocols on both accuracy and communication efficiency, and in some cases outperformed a baseline where every agent could see all the evidence upfront.
Multi-agent LLM setups are creeping into real products, from coding assistants to research tools, but most just let models talk in a fixed order or a free-for-all, with no way to check whether any given exchange was useful. Consilience's contribution isn't a smarter model. It's a statistically certified process for deciding who should talk and how, which is the unglamorous infrastructure work that determines whether these systems are reliable enough to trust.
The more striking result is buried in the numbers: better-managed conversation beat simply handing every agent all the information. Worth remembering next time a pitch deck promises answers by just throwing more context at the problem.