Codebook Agent: Amortized Topology Design for LLM Multi-Agent Systems

Summary

The current state of the art in automated topology search, and useful to the paper for two negative findings rather than its method: topologies that survive a reward filter collapse to roughly six distinct graphs even as codebook capacity grows from 8 to 64, and edge count is negatively correlated with measured token consumption so sparsifying a graph makes inference more expensive. It also notes that message-passing scorers over agent-profile nodes are adjacency-invariant when agents share a profile, the default in published benchmarks, meaning existing topology-search literature largely cannot separate structure from role heterogeneity. That is a direct argument for the paper's deliberate crossing of structure with aggregation rule instead of black-box search.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a medium-priority candidate for the Architecture as Epistemology (ISR) paper (angle: MAS topology / structure vs. aggregation rule). Verified against: Opened https://arxiv.org/abs/2609.02264 and read the full abstract including the six-graph collapse and Pearson r about -0.4 findings; arXiv API confirms nine authors, published 2026-09-02, cs.AI.

Critique / open questions

Queued — not yet assessed.

Key quotes

Adapting the communication topology of an LLM multi-agent system to each query improves both accuracy and efficiency, yet current designers treat this as conditional graph generation: a variational, autoregressive, or diffusion decoder searches the $N \times N$ adjacency space, and a graph-network proxy trained on utility and a structural cost such as edge count ranks the sampled candidates. (abstract)

We argue that this formulation is misaligned with the problem. (abstract)