What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems
Original reporting by arXiv (cs.AI)
Multi-agent systems (MAS) powered by large language models are transforming AI capabilities, yet their internal communication often hides a significant efficiency bottleneck. While agents are carefully assigned roles and tasks, the information they exchange with one another typically relies on unconstrained natural language. This seemingly intuitive approach, however, comes at a substantial cost: verbose exchanges rapidly consume precious context windows, inflate token usage, and ultimately slow down performance while driving up inference expenses. It's a fundamental challenge to scaling complex AI collaboration effectively.
A new communication protocol
Researchers delved into this dilemma, analyzing five common inter-agent communication strategies across various system architectures. Their findings revealed that no single fixed approach was universally optimal. Instead, effective communication consistently distilled messages down to the "action-centered information" critical for downstream agents. Building on this insight, they introduced PACT (Protocolized Action-state Communication and Transmission). PACT reframes inter-agent dialogue as a public state-update problem, compacting raw agent outputs into concise action-state records before they enter shared history. This novel protocol consistently enhances the performance-cost trade-off, achieving comparable or superior task results with substantially fewer tokens, a benefit demonstrated even in production coding environments like OpenHands and SWE-agent. PACT offers a blueprint for more efficient and scalable AI team dynamics.
The proliferation of multi-agent systems built upon large language models has underscored a critical need for efficient inter-agent communication. PACT (Protocolized Action-state Communication and Transmission) directly addresses this challenge by meticulously refining how agents share information. Instead of relying on unconstrained natural language, PACT projects raw agent outputs into concise, action-state records, dramatically curbing token usage without compromising performance. This innovative approach consistently yields superior performance-cost trade-offs across various MAS topologies, evidenced by its ability to boost OpenHands' resolve rate with 10% fewer tokens and halve input tokens for SWE-agent while maintaining task efficacy. PACT thus offers a compelling framework for building more economical and performant current-generation agent systems.
Forging Efficient AI
Looking beyond immediate gains, PACT’s implications for the broader landscape of artificial intelligence are substantial. By making multi-agent systems inherently more resource-efficient and predictable, PACT lowers the operational barriers to deploying complex AI solutions in real-world scenarios. This newfound efficiency does not merely optimize existing systems; it fundamentally expands the possibilities for constructing more sophisticated, robust, and scalable AI architectures. As the demand for intelligent automation grows, the ability to manage the computational overhead of interacting agents will become a defining factor in AI adoption. PACT represents a pivotal step towards a future where AI systems are not only intelligent but also sustainably efficient, enabling a new era of collaborative and impactful artificial intelligence.