PACT protocol slashes AI agent communication costs by 50%

Multi-agent systems (MAS) built on large language models typically let agents communicate in unconstrained natural language, but this free-form chit-chat rapidly inflates token usage and consumes the shared context window, harming both system performance and inference cost. In a new arXiv preprint, researchers Chen Huang, Yuhao Wu, and Wenxuan Zhang analyze five common inter-agent communication strategies across two MAS topologies. Their key finding: no fixed strategy is universally optimal, but effective messages consistently preserve action-centered information needed by downstream agents. Building on this insight, they propose PACT (Protocolized Action-state Communication and Transmission), which treats inter-agent communication as a public state-update problem and projects each raw agent output into a compact action-state record before entering shared history.

In extensive experiments, PACT consistently improves the performance-cost trade-off across different MAS topologies, achieving comparable or stronger task performance with substantially fewer tokens. The gains extend to production coding harnesses: PACT lifts OpenHands' resolve rate while using 10% fewer tokens per resolved task, and is resolve-neutral on SWE-agent while halving input tokens. The authors have released their code publicly. This work offers a practical protocol for building more efficient and cost-effective AI agent systems, especially relevant as multi-agent deployments grow in enterprise and research settings.