BANDMAS cuts multi-agent LLM traffic by 77% with smart packet scheduling

LLM-based multi-agent systems exchange large volumes of messages to make decisions, but every transmitted token consumes expensive context windows and adds inference latency. Existing fixes—pruning agents or discarding redundant messages—often strip away critical evidence or still leave wasteful transmissions. BANDMAS, a new framework from Jiangwen Dong and Wanyu Lin, tackles this at the packet level. It models inter-agent communication as task-oriented traffic and decomposes each message into semantic packets tagged as evidence or requests. A causality-inspired replay valuation predicts whether a packet's eventual contribution to the final decision justifies its bandwidth and compute cost. The system then schedules packets adaptively under hard constraints like bandwidth, latency, deadline, and receiver context.

In experiments using frozen Qwen3-4B traffic on SciFact, HotpotQA, and FanOutQA benchmarks, BANDMAS reduced application-layer bytes by 53.2% to 77.3% at selected caps, while achieving the highest mean task metric among constrained methods across all three workloads. This suggests that selectively dropping low-value packets—instead of whole agents or messages—strikes a better balance between communication efficiency and decision quality. For enterprises running multi-agent pipelines, BANDMAS could materially lower token costs and response times, especially in bandwidth-limited or latency-sensitive settings. The research is available under arXiv:2608.00458, and while it's still an academic proposal, its packet-level semantics approach points toward a more principled future for agent orchestration.