MACA framework boosts multi-agent AI by 8.42% using 43% fewer tokens

A team of researchers (Haoran Li, Shulun Chen, Shaoyuan Sun, Hanchen Wang) has released a preprint on arXiv introducing MACA (Multi-Agent Coordination Adaptation via Structure-Guided Orchestration), a new framework to address a fundamental tension in LLM-based multi-agent systems: the trade-off between structural stability and dynamic adaptability. Existing approaches either fix the agent coordination structure upfront (structure-centric) or adapt decisions on the fly without explicit coordination (orchestration-centric). MACA reframes the problem as posterior inference over the joint distribution of structure and orchestration, enabling both fine-grained control and dynamic flexibility.

The framework first learns a task- and budget-conditioned structural prior that dictates which agents participate and how they interact. This prior then guides a policy-based orchestration engine that approximates optimal posterior inference. In experiments, MACA outperformed adaptive multi-agent baselines by an average of 8.42% while consuming 43.19% fewer tokens. Further analysis shows the joint adaptation suppresses redundant agent interactions, converging coordination toward task-effective execution. The paper is 21 pages and available on arXiv (2605.25746), with implications for scaling complex agentic workflows in production AI systems.