What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems

Researchers propose PACT, a new protocol for multi-agent AI systems that compresses inter-agent communication into compact action-state records, reducing token usage by up to 50% while maintaining or improving task performance. The approach addresses a critical efficiency bottleneck in large language model-based multi-agent systems, with demonstrated improvements in production coding applications.

Multi-agent systems powered by large language models face a fundamental efficiency challenge: agents communicating in unconstrained natural language rapidly consume token budgets and shared context windows, inflating inference costs while potentially degrading performance. This research identifies that effective inter-agent communication doesn't require preserving all conversational details—only action-centered information that downstream agents actually need to execute their tasks. The PACT protocol formalizes this insight by treating agent communication as a public state-update problem, converting raw agent outputs into standardized, compact records before storing them in shared history.

The work is timely given the explosive growth of multi-agent frameworks in production environments. As organizations deploy increasingly complex reasoning systems, token efficiency directly impacts operational costs and feasibility. Traditional approaches either accept bloated communication or risk losing information critical for coordination. PACT achieves a genuine breakthrough by demonstrating that information loss through compression is often recoverable through better structure.

The empirical results carry real implications for AI development teams. OpenHands achieved 10% token reduction while improving resolve rates, while SWE-agent halved input tokens without performance degradation. These production-level results, beyond controlled research settings, suggest PACT addresses genuine infrastructure pain points. The approach also generalizes across different multi-agent topologies, indicating broad applicability rather than optimization for specific scenarios.

Developers building multi-agent systems should monitor whether PACT becomes standard practice. The framework's public code release enables rapid adoption. Longer-term, this work signals that AI system efficiency improvements may come from better communication protocols rather than model scaling, potentially affecting how organizations architect and cost their agent infrastructure.

• →PACT reduces token usage in multi-agent systems by up to 50% while maintaining or improving task performance through structured action-state communication.
• →The protocol works across different multi-agent topologies and demonstrates effectiveness in production systems like OpenHands and SWE-agent.
• →Effective inter-agent communication requires preserving only action-centered information, not full conversational context.
• →Token efficiency improvements from better protocols may reduce AI system operational costs without requiring model upgrades.
• →The approach treats agent communication as a standardized state-update problem rather than unconstrained natural language exchange.