🤔 Does the communication topology for LLM multi-agent systems really need to be generated from scratch every single time? A team at UCLA questioned that very assumption. Title: Codebook Agent: Amortized Topology Design for LLM Multi-Agent Systems URL: ❓ Why have people assumed agent communication topologies need to be searched by a generative model every time? 💡 Turns out they may not need to be. The topologies that survive reward filtering converge to only about 6 distinct graphs, even as the codebook size grows from 8 to 64 — the design space is far smaller than it looks. ❓ Wouldn't making the graph sparser (fewer edges) cut down token usage? 💡 The opposite happens. Edge count correlates with token cost at r=-0.4 — sparser graphs actually use more tokens. Structural "cheapness" doesn't match measured cost. ❓ What's wrong with scoring candidate topologies using a GNN, as prior work does? 💡 For homogeneous agent teams (common in real deployments), the GNN's message passing treats every adjacency matrix as the same input, so it can't tell candidates apart — a real failure mode, not just a minor gap. ❓ So how well does Codebook Agent actually work once it swaps generation for selection? 💡 By compressing topologies into 16 codes via a VQ-AE and picking candidates with a reward-weighted MLP plus a proxy trained on real measured costs, it cuts generation latency from 301-396ms down to 2.4ms — a 125-158x speedup — while improving accuracy by 1.6 points on average over the prior best method and cutting tokens by 21.9-33.2%. #MultiAgent# #LLM#