The Big Picture
Effective communication layouts for teams of large language model agents fall into a tiny set of reusable patterns; picking from that codebook with a lightweight scorer matches or beats costly search while cutting design time to 2.4 ms and saving ~22–33% of tokens.
ON THIS PAGE
The Evidence
**WHAT THEY FOUND:** Successful team communication graphs cluster into about six distinct patterns regardless of how large the model used to design them is, so generating full adjacency matrices is unnecessary. Capability Discovery Pattern Counting edges is a poor proxy for run-time cost—sparser graphs can actually produce longer (and more expensive) outputs, so minimizing edge count can increase cost. A simple offline codebook (learned from past runs), a fast query-to-code predictor, and a single batched proxy that scores decoded candidates match or beat heavier iterative designers on accuracy while drastically reducing latency and token use.
Data Highlights
1Topologies compress to roughly six effective codes even as codebook capacity grows (used codes saturate near 6 for K≥8).
2Topology generation time dropped to 2.4 milliseconds with no sampling loop or message-passing at test time.
3Measured LLM token use fell by 21.9%–33.2% across benchmarks compared to competing methods.
What This Means
**WHO SHOULD CARE:** Engineers building multi-agent LLM systems who need faster per-query adaptation and lower token bills can adopt a codebook-style selector to get near-best accuracy with much lower overhead. Multi-Agent System Technical leaders and infra teams responsible for production costs and latency will find this useful for scaling agent orchestration without expensive online search. Researchers studying agent interaction can use the finding that effective designs form a short list to simplify future evaluation and design strategies.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1: Three regimes of topology design. (Top) A fixed topology is reused for every query: zero design cost, but no single graph fits every task. (Middle) Query-conditioned generators adapt the graph via an iterative loop that samples and scores K K candidates at each of T T steps, typically with a proxy trained on utility and edge count. (Bottom) Codebook Agent keeps per-query adaptation and removes the loop: the design space is discretized once offline into a codebook, a query embedding selects codes, and one batched proxy call under a token-based cost objective returns the winner.
Fig 2: Figure 2: Overview of Codebook Agent. Left: offline collection executes fixed topologies on the training split and logs ( A , c , u , τ ) (A,c,u,\tau) for every (query, topology) pair. Middle: (1) a VQ autoencoder compresses successful topologies ( u > 0.5 u>0.5 ) into a query-independent codebook; (2) an MLP predictor p θ ( k ∣ c ) p_{\theta}(k\mid c) is fit to the reward-weighted soft targets of Eq. ( 6 ); (3) an MLP proxy f ϕ f_{\phi} is regressed on measured utility and per-task normalized token cost (Eqs. ( 8 )–( 10 )), with an auxiliary structure-token head used in training only. Right: at test time the top M M codes are decoded, deduplicated, and scored in one batched proxy call; the candidate maximizing u ^ − λ c ^ \hat{u}-\lambda\hat{c} (Eq. ( 11 )) is executed. Dashed arrows are training-only; solid arrows are the frozen test-time path (no LLM calls inside topology generation).
Fig 5: Figure 5: Ablations on GSM8K and homogeneous HumanEval. (a) Accuracy versus team size N N (retrained per N N on the same training budget). (b) Accuracy versus codebook size K K ; used codes saturate near six for every K ≥ 8 K{\geq}8 . (c) Mean tokens per query for MLP, random, and GNN reranking under a fixed candidate set. (d) Mean tokens versus the rerank cost weight λ cost \lambda_{\mathrm{cost}} , with default − 0.1 -0.1 .
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
**CONSIDERATIONS:** The codebook is trained from measured runs, so it depends on having an execution log of topologies and costs; cold-starts or novel task families may require new collection. execution log Most experiments use homogeneous agent teams and a fixed set of benchmarks—performance on highly heterogeneous teams or radically different role definitions may differ. The token-cost proxy requires measured token counts during training; if your backend or billing model changes, the proxy should be retrained or recalibrated.
Methodology & More
**FULL SUMMARY:** Collect execution records of agents under several fixed graphs, then compress the successful graphs into a small discrete codebook using a vector-quantized encoder. Train a fast predictor that maps a query embedding to a distribution over those codes, and train a lightweight multi-layer network to predict measured utility and normalized token cost from a decoded adjacency. At test time, decode the top candidate codes, deduplicate them, score them in one batched proxy call, and execute the highest-scoring graph—no iterative sampling or graph message passing required. Across six reasoning and coding benchmarks and two backbone models, this amortized approach equals or outperforms iterative topology generators on accuracy, while shrinking per-query design latency to about 2.4 ms and cutting LLM token consumption by roughly 22–33%. Key practical takeaways are that the useful design space is small, edge count is a misleading cost proxy, and graph neural critics can be blind on homogeneous teams—so a small, measured, and execution-grounded toolkit is often the simplest, fastest, and cheapest solution for multi-agent orchestration. consensus-based decision pattern Agentic RAG Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Most listed h-indexes are very low (≈3) and affiliations are unspecified. No clear senior/top-lab authors or top conference venue—fits emerging/limited-info category.