Codebook Agent Optimizes LLM Multi-Agent Communication Topology

Optimizing the communication topology within a multi-agent system composed of Large Language Models (LLMs) is crucial for enhancing both accuracy and efficiency for specific queries. Current methods often treat this as a conditional graph generation problem, involving complex variational, autoregressive, or diffusion decoders to search a vast adjacency space, with graph-network proxies ranking sampled candidates based on utility and structural costs like edge count.This research argues that this traditional formulation is misaligned with the actual problem. Empirical observations show that effective topologies often collapse into a small number of distinct graphs, even with increased codebook capacity. Counter-intuitively, sparsifying the graph (reducing edge count) can actually increase token consumption, making inference more expensive. Furthermore, message-passing scorers often become adjacency-invariant when agents share profiles, which is common in benchmarks, rendering them ineffective for ranking.To address these issues, the paper introduces Codebook Agent. This framework employs a vector-quantized autoencoder to compress successful communication topologies into a compact, query-independent 16-entry codebook. A reward-weighted Multi-Layer Perceptron (MLP) then maps the query embedding to a distribution over these codes. Finally, another MLP proxy, trained on measured utility and per-task normalized token cost, reranks the top decoded candidates in a single batched forward pass. This approach eliminates iterative search and message passing at test time, making Codebook Agent the most accurate method across six benchmarks (84.6% average vs. 83.0% for prior designers), while generating a topology in just 2.4 ms and using 21.9-33.2% fewer LLM tokens.