Adaptive Multi-Agent Topology Collapses to Six Graphs, and Sparser Graphs Cost More Tokens

arXiv 2609.02264·medium signalXBlueskyLinkedInCopy link

Codebook Agent argues the standard framing of per-query multi-agent topology design as conditional graph generation over the N×N adjacency space is misaligned with the problem, and backs it with three empirical facts. Topologies surviving a reward filter collapse to about six distinct graphs even as codebook capacity grows from 8 to 64; edge count is negatively correlated with measured token consumption (Pearson r ≈ -0.4), so sparsifying the graph makes inference more expensive, not cheaper; and a message-passing scorer over agent-profile nodes is adjacency-invariant whenever agents share a profile, the default in published benchmarks, so it cannot rank candidates at all there. Their replacement uses a VQ autoencoder compressing successful topologies into a query-independent 16-entry codebook plus an MLP proxy reranking decoded candidates in one batched forward pass, with no iterative search or test-time message passing.