Published on June 6, 2026
Multi-agent systems (MAS) using large language models once relied on free-form communication, which is natural but often inefficient. Agents were organized schedules, trading lengthy messages that led to high token usage and slow response times. This method, while adaptable, became a bottleneck in performance and cost.
Recent research has identified the inefficiencies caused . In analyzing various communication strategies in MAS, it emerged that no single method worked universally. Instead, the researchers found that messages which focus on key action-centered information yield better outcomes for downstream agents.
To address these challenges, the concept of PACT (Protocolized Action-state Communication and Transmission) was introduced. This innovative approach standardizes communication agent’s output into a compact action-state record. The researchers demonstrated that PACT not only enhances the system’s efficiency but also preserves vital information without burdening the shared context.
The impact of PACT is significant across multiple MAS frameworks. Testing showed that it improves the performance-cost ratio, achieving strong task performance with a reduced token footprint. In practical applications such as OpenHands and SWE-agent, PACT maintained effectiveness while halving input tokens, suggesting a promising direction for future developments in agent-based communication.
Related News
- Keyence Shares Soar 16% After Strong Earnings and Buyback Announcement
- Major Data Breach Hits NYC's Largest Public Healthcare System
- Korea's President Addresses Controversy Over AI Profit Redistribution
- Amazon Reduces Workforce and Discontinues Grocery Service in Singapore
- Pope and AI Firm's Unlikely Alliance Raises Eyebrows
- AI-Powered Toys Challenge Parenting Norms