Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems

1 INTRODUCTION

• Existing communication methods consume 2–11.8× as many tokens as the simple chain structure, adding economic burdens that challenge large-scale deployment.The introduction notes that excessively costly inference is also unsuitable for edge-device deployment.Source: §1 · p. 2
• Communication redundancy occurs when substantial message passing does not meaningfully contribute to collaborative intelligence.The paper models agents as graph nodes, within-dialogue communication as spatial edges, and cross-dialogue communication as temporal edges.Source: §1 · pp. 2–3
• AgentPrune trains a low-rank-principle-guided graph mask to identify key exchanges, then prunes non-essential communication in one shot.The resulting sparse graph is used as a fixed topology for subsequent reasoning; the low-rank principle is also used to robustify systems against agent-targeted attacks.Source: §1 · p. 3
• At $5.6 cost, AgentPrune achieves comparable MMLU performance to state-of-the-art communication topologies costing $43.7.Across six benchmarks, it also reduces token costs by 28.1%–72.8% in existing frameworks and improves performance against two types of agent adversarial attacks by 3.5%–10.8%.Source: §1 · p. 3

2 LLM-MA AS SPATIAL-TEMPORAL GRAPHS

• Each graph node represents an agent; spatial edges carry information within an utterance, while temporal edges pass messages between dialogue rounds.Agent nodes include a language model, role, interaction state, and available plugins.Source: §2 · p. 3
• The spatial communication graph must be a directed acyclic graph, so agents can be processed in topological order after their dependencies.An agent produces a rationale or answer from the query, its role and state, incoming messages, and prompting instructions; after K rounds, aggregation yields the final solution.Source: §2 · p. 3
• Communication redundancy exists when a subgraph using subsets of spatial and temporal edges achieves solution quality at least as high as the full graph.The redundant components are the removed spatial and temporal edges, with solution quality measured by a utility function.Source: §2 · p. 4
• The study objective is to minimize communication redundancy while keeping performance variation within an allowable threshold.The supplied passage describes Equation (6) as minimizing redundancy with a performance guarantee.Source: §2 · p. 4

3 METHODOLOGY

• Figure 4’s overview proceeds from spatial pruning within a dialogue round to temporal pruning of dialogue history.The method then derives a sparse interaction topology through one-shot pruning.Source: §3 · p. 4, Figure 4 · p. 5

3.1 SPATIAL-TEMPORAL GRAPH COMMUNICATION

• AgentPrune relaxes a system’s spatial and temporal adjacency matrices with differentiable masks, while DAGSampling keeps the spatial graph acyclic.The input system supplies the predefined adjacency matrices.Source: §3.1 · p. 5

3.2 OPTIMIZING SPATIAL-TEMPORAL CONNECTIVITY

• The masks are optimized for utility and token efficiency using distribution approximation and low-rank sparsity objectives.Because utility evaluation is often non-differentiable, the method uses policy gradients; it replaces NP-hard rank minimization with the nuclear norm.Source: §3.2 · pp. 5–6

3.3 ONE-SHOT PRUNING

• After limited mask optimization, AgentPrune performs one-shot magnitude pruning and fixes the resulting sparse connectivity for the remaining rounds.The pruning retains the largest mask entries according to the selected pruning ratio, then applies binary masks to the original topology.Source: §3.3 · p. 6

3.4 APPLICATION AND ANALYSIS

• AgentPrune is a plug-and-play module for systems with more than three agents and moderately organized communication, and can train topology using Q′≪Q queries.It uses K′ interaction rounds with trainable masks before pruning, then applies the sparse topology for the remaining K−K′ rounds.Source: §3.4 · pp. 6–7

4 EXPERIMENTS

• The experiments ask whether AgentPrune improves task completion and token efficiency, reduces framework costs without compromising performance, resists agent attacks, and responds to component or parameter changes.Source: §4 · p. 7

4.1 EXPERIMENTAL SETUP

• Experiments cover general reasoning, mathematical reasoning, and code generation, comparing single-agent, spatial, and temporal communication baselines across GPT-3.5 and GPT-4 settings.The setup uses dialogue rounds K=2 for general and mathematical reasoning and K=4 for code generation; multi-query settings use Q′ in {5, 10}.Source: §4.1 · pp. 7–8

4.2 PERFORMANCE & COST COMPARISON (RQ1)

• 90.3% on HumanEval and 95.8% on GSM8K, AgentPrune-R uses less than 40% of DyLAN’s token consumption on both benchmarks.Figure 5 plots performance against prompt-token consumption across MMLU, HumanEval, and GSM8K.Source: §4.2 · p. 8, Figure 5 · p. 8

4.3 PLUG-IN INTO EXISTING FRAMEWORKS (RQ2)

• 60.6% fewer prompt tokens and up to $177.58 saved on GSM8K+GPTSwarm, with performance increasing by 0.84%.For HumanEval+AutoGen, AgentPrune reduces prompt tokens by 36% and saves $1.486.Source: §4.3 · p. 9

4.4 ROBUSTNESS VERIFICATION (RQ3)

• 78.4% to 83.9% under agent prompt attacks, complete-graph accuracy rises with AgentPrune, while un-attacked performance rises from 83.1% to 84.7%.Across attacks, AgentPrune boosts DyLAN and AutoGen robustness by up to 6.3%; its impact on GPTSwarm is marginal given that framework’s inherent defenses.Source: §4.4 · pp. 9–10, Figure 6 · p. 9

4.5 EXPERIMENTAL ANALYSIS

• Agent profiling has variable utility across datasets, with more pronounced effects on general reasoning and code generation than math reasoning; low-rank sparsity consistently aids topology optimization.The ablation study is reported in Table 6 and Appendix H.4.1.Source: §4.5 · p. 10

• Intra-dialogue communication covers message exchange within a dialogue round, while inter-dialogue communication passes information between rounds through full transmission, partial transmission, or summarization.Source: §5 · p. 10
• Earlier multi-agent graph learning uses graph diffusion, weighted GNNs, or transformers, while LLM-agent systems use graphs to model interactions or jointly optimize prompts and inference structure.Source: §5 · p. 10
• Existing LLM-agent graph approaches do not jointly optimize intra- and inter-dialogue structures and often increase token consumption.Source: §5 · p. 10

6 CONCLUSION

• AgentPrune is presented as an economical, simple, and robust communication pipeline that embeds into mainstream multi-agent frameworks and prunes communication redundancy.Source: §6 · p. 10
• The paper concludes that AgentPrune achieves performance comparable to or better than original systems with significantly smaller token throughput and economic costs.Source: §6 · p. 10
• The authors suggest AgentPrune can facilitate progress toward larger-scale collective intelligence.Source: §6 · p. 10

• For complex tasks requiring more than five dialogue turns, the text describes optimizing topology over K′ in {1, 2} rounds before continuing for K − K′ rounds.Source: D MULTI-QUERY TRAINING OF AgentPrune · p. 16
• For benchmarks with Q queries, the first Q′, where Q′ << Q, serve as a training phase to optimize the spatio-temporal topology; a one-shot pruning then fixes it for the remaining Q − Q′ queries.Source: D MULTI-QUERY TRAINING OF AgentPrune · p. 17
• Query-level optimization uses early benchmark queries to inform a topology for the remaining queries, unlike dialogue-level optimization, which uses initial utterances while resolving one query.Source: D MULTI-QUERY TRAINING OF AgentPrune · p. 17

E COST ANALYSIS

• AgentPrune optimizes connectivity over the first Q′ queries, then uses the pruned graph for the remaining Q − Q′ queries.The cost analysis assumes K dialogue rounds per query and denotes average spatial, temporal, and query-message token counts as cS, cT, and cq.Source: E COST ANALYSIS · pp. 17–18

F.1 CHAIN STRUCTURE

• In a chain, the first agent transforms the user’s instruction and passes it onward until the final agent provides a solution.The passage cites ChatDev, MetaGPT, and L2MAC as examples.Source: §F.1 · p. 18, Figure 7 · p. 19

F.2 TREE STRUCTURE

• The tree topology is presented as a spatial communication structure.Figure 8 demonstrates the tree structure.Source: Figure 8 · p. 19

F.3 STAR STRUCTURE

• The star topology is presented as a spatial communication structure.Figure 9 demonstrates the star structure.Source: Figure 9 · p. 19

F.4 COMPLETE GRAPH STRUCTURE

• The paper’s “complete graph” is a directed DAG that would be complete if made undirected; topological ordering lets agents execute in sequence.The final agent summarizes the dialogue and provides a concluding output or reflection.Source: §F.4 · p. 20, Figure 10 · p. 20

F.5 LAYERED GRAPH STRUCTURE

• The layered topology is presented as a spatial communication structure.Figure 11 demonstrates the layered graph structure.Source: Figure 11 · p. 20

F.6 RANDOM GRAPH STRUCTURE

• A random graph is sparsely sampled from a complete graph; its random shortcuts are associated with shorter average paths and small-world characteristics.The passage attributes reported advantages over regular fully connected structures to random edge connections.Source: §F.6 · p. 20, Figure 12 · p. 21

F.7 LLM-BLENDER STRUCTURE

• LLM-Blender is presented as a spatial communication topology.Figure 13 demonstrates its structure.Source: Figure 13 · p. 21, §F.7 · p. 21

G.1.1 SPATIAL COMMUNICATION BASELINES

• Spatial baselines run for two rounds without explicit inter-dialogue messaging unless stated otherwise; chain, tree, star, complete, layered, and random structures use distinct message-routing arrangements.The complete and random graph implementations use DAG sampling, topological execution, and a final summarizer; the layered graph’s final agent receives outputs from the preceding layer.Source: §G.1.1 · pp. 21–22
• LLM-Blender uses GenFuser’s output and single-turn dialogues, while GPTSwarm uses agent profiles and was modified to transmit both choices and reasoning on MMLU.The evaluation standardizes LLM-Blender agents to gpt-3.5 or gpt-4.Source: §G.1.1 · pp. 22–23

G.1.2 TEMPORAL COMMUNICATION BASELINES

• LLM-Debate shares all agents’ prior-round outputs with every agent; PHP uses each round’s decisions as hints, and DyLAN removes low-quality outputs between rounds.The LLM-Debate and PHP structures are shown in Figures 14 and 15.Source: §G.1.2 · p. 23, Figure 14 · p. 23, Figure 15 · p. 23

G.1.3 OTHERS

• AutoGen is evaluated with five roles or a condensed three-role configuration, using its Dynamic Group Chat setup.The five roles are user proxy, manager, engineer, critic, and code executor; the three-agent version retains manager, engineer, and critic.Source: §G.1.3 · p. 24, Figure 16 · p. 24

G.2 AGENT PROFILING

• The authors used gpt-4 to generate task-specific agent profiles, drawing on prior work that associates distinct personas with cognitive synergy.The appendix gives general reasoning, mathematical reasoning, and code-generation profile examples.Source: §G.2 · p. 24, §G.2.1 · p. 24, §G.2.2 · p. 25

G.2.1 PROFILE EXAMPLES FOR GENERAL REASONING

• Math profiles ask agents to solve from hints, analyze variables before substitution, or combine reasoning with Python code.General-reasoning and code-generation examples include roles such as entity researcher, critic, project manager, algorithm designer, programmer, and bug fixer.Source: §G.2.1 · p. 24, §G.2.2 · pp. 25–26

G.3 AGENT ATTACK IMPLEMENTATION

• The evaluation implements prompt attacks by changing an agent’s role to a liar and replacement attacks by substituting a random-text dummy; one agent is selected per framework.The remaining agents retain their original functions and responsibilities.Source: §G.3 · p. 27

H.1 RESULTS FOR RQ1

• On MMLU, AgentPrune surpasses GPTSwarm at $5.6 versus $43.56; on GSM8K, it outperforms DyLAN at $65.9 versus $357.47.Figures 17–19 compare completion tokens, overall tokens, and overall cost across datasets.Source: Figure 19 · p. 28, Figure 17 · p. 28, Figure 18 · p. 28

H.3 RESULTS FOR RQ3

• Figure 20 compares framework performance before and after agent replacement attacks, with “w/ AP” marking AgentPrune integration.The supplied passage describes the comparison but does not report its outcome.Source: Figure 20 · p. 29, §H.3 · p. 29

H.2 RESULTS FOR RQ2

• Table 5 compares performance, prompt and completion token use, and cost when AgentPrune is integrated with three-agent AutoGen and GPTSwarm.The table caption specifies a three-gpt-4-based setting.Source: §H.2 · p. 29, Table 5 · p. 29

H.4.1 ABLATION STUDY

• The ablation study compares variants without unique agent profiles and without low-rank regularization.Table 6 labels these variants “w/o profile” and “w/o low-rank.”Source: §H.4.1 · p. 29, Table 5 · p. 29

H.4.2 SENSITIVITY ANALYSIS

• Accuracy improves significantly from 3 to 5 agents, while gains from 5 to 9 agents are relatively marginal.Figure 21 varies |V| from 3 to 9 agents on MMLU.Source: §H.4.2 · p. 29, Figure 21 · p. 30
• The early stopping round Q′ is the number of queries used to optimize connectivity; the authors set Q′ ∈ {5, 10} to balance performance with token savings.Increasing optimization rounds is described as refining the graph mask and reducing performance fluctuations.Source: §H.4.2 · p. 30, Figure 21 · p. 30

I.1 SPATIAL TOPOLOGY VISUALIZATION

• The pruned spatial graph may not be a Directed Acyclic Graph, so DAGSampling is still applied before agents execute sequential I/O.The passage identifies DAGSampling as the step that ensures the required DAG properties.Source: §I.1 · p. 30
• In five-agent GPTSwarm on MMLU, AgentPrune removes many incoming edges from simple I/O agents while preserving many for Tree of Thoughts agents.The passage suggests stronger reasoning capabilities may make Tree of Thoughts agents better suited to synthesize discussion and provide the final solution.Source: §I.1 · p. 30, Figure 22 · p. 31
• On MMLU role graphs, the Critic retains many incoming edges and the Knowledge Expert many outgoing edges; the psychologist has no outgoing edges in the random-graph example.The authors say the psychologist's missing outgoing edges may suggest limited utility in the overall system.Source: §I.1 · p. 31, Figure 24 · p. 32, Figure 25 · p. 32
• In the HumanEval code-completion example, outer edges are retained and the Bug Fixer has no outgoing edges as the workflow's final step.The passage describes this as illustrating autonomous optimization of collaborative topologies.Source: §I.1 · p. 31, Figure 26 · p. 33
• In the four-role math example, analysis agents have high out-degree and no incoming edges, while problem-solving agents have high in-degree and low out-degree.A separate two-solver setup differentiates preliminary solving from final solving; the final solver gathers information from all other nodes and has no outgoing edges.Source: §I.1 · p. 32, Figure 28 · p. 34

I.2 TEMPORAL TOPOLOGY VISUALIZATION

• Figures 30–31 show AgentPrune operating on predefined temporal communication structures and resulting structures after one-shot pruning; Figures 32–34 show spatial connectivity masks differentiating over time before pruning, with magnitudes informing subsequent pruning decisions.Source: §I.2 · p. 33, §I.3 · p. 33