AGTBRIEF
A new preprint trains a 'reward model' to keep teams of AI agents from over chatting, trimming 20.5% off their token bill. The saving is the paper's own measurement.
When a team of AI agents breaks a hard problem into steps and trades messages to solve it, every round-trip costs money, and the bill scales with how much the agents say to each other. A new arXiv preprint reports a way to trim that cost, claiming an average 20.5% cut in tokens while preserving the team's accuracy on the authors' own benchmark.
The method, called RGA-Designer, treats the conversation between agents as a graph, a map of who talks to whom. Its predecessor, ARG-Designer, learned to generate those graphs one node at a time but offered no training signal for keeping them lean. The new work borrows from reinforcement learning from human feedback: train a small reward model to score not just whether the final answer is correct, but whether the graph itself is sparse and efficient, then fine-tune the graph generator against that score.
The catch: the 20.5% saving is the authors' own average across their evaluation suite. Token cost is one lever among several, sitting next to prompt caching, smaller specialist models, and router patterns. The paper has not been peer-reviewed or independently replicated.