AgentPrune cuts multi-agent LLM communication costs from $43.7 to $5.6 in a new paper

sebkrier · x · 2026-07-27

A new paper proposes AgentPrune, a communication pipeline for LLM-based multi-agent systems that prunes redundant or malicious messages.

The paper argues that existing multi-agent setups waste tokens and money because communication topologies are carefully designed but often overly verbose.
AgentPrune performs one-shot pruning on the spatial-temporal message-passing graph to remove redundant communication while preserving performance.
In experiments across six benchmarks, the authors claim it matches state-of-the-art topologies at $5.6 vs $43.7, cuts tokens by 28.1%–72.8%, and improves robustness against two adversarial attack types by 3.5%–10.8%.
The source code is available on GitHub: yanweiyue/AgentPrune.

More from coding & agent