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

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Advanced, extended

Summary

Multi-agent systems (MAS) built on large language models often suffer from high token usage and context window exhaustion due to unconstrained natural language communication. A diagnostic analysis of five strategies across two MAS topologies (split-evidence interaction and sequential pipeline) found no universally optimal approach, but highlighted the need for action-centered information. Researchers propose PACT (Protocolized Action-state Communication and Transmission), a protocol that projects raw agent output into a compact "action-state record" (Action, State, Result) for shared history, separating private computation from public communication. PACT consistently improves the performance–cost trade-off, reducing token usage by 38.7% on average. It boosts OpenHands' resolve rate by 18 instances with a 10.3% tokens-per-resolved decrease and halves SWE-agent's input tokens (50.4% reduction) while maintaining resolve rates. The code is publicly available at https://github.com/iNLP-Lab/PACT.

Key takeaway

For AI Engineers developing multi-agent systems, unconstrained natural language communication significantly inflates token costs and can degrade performance. You should implement a structured communication protocol like PACT, which distills agent outputs into compact action-state records. This approach reduces token usage by up to 50.4% and maintains or improves task performance, ensuring your MAS remains efficient and effective, especially in long-context applications like agentic coding harnesses.

Key insights

Structured, action-state communication significantly reduces token usage and improves multi-agent system efficiency.

Principles

• No single communication strategy is universally optimal for MAS.
• Inter-agent messages must be action-centered and grounded.
• Separate private agent computation from public communication.

Method

PACT projects raw agent output into a compact, three-field action-state record (Action, State, Result) before appending it to the shared history, thereby excluding intermediate process content.

In practice

• Use a proxy hook to filter agent outputs in-flight.
• Define explicit Action, State, and Result fields.
• Retain tool calls and results, but strip verbose prose.

Topics

• Multi-agent Systems
• LLM Communication
• Token Efficiency
• PACT Protocol
• Agentic Coding
• Context Management

Code references

• iNLP-Lab/PACT

Best for: NLP Engineer, AI Architect, Research Scientist, AI Scientist, AI Engineer, Machine Learning Engineer

Related on AIssential

• What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems — Unconstrained natural language communication in MAS inflates costs; structured action-state records improve efficiency and performance.
• When AI Talks to AI: A Practical Guide to Agent-to-Agent Communication — Distributing complex AI tasks across multiple communicating agents overcomes single-agent limitations.
• Beyond tokens: a unified framework for latent communication in LLM-based multi-agent systems — Latent communication in LLM multi-agent systems reduces cost and information loss by exchanging continuous representations directly.
• Context Engineering for AI Agents: A Deep Dive — Precise context shaping, not just quantity, is critical for robust multi-agent system performance and reliability.
• Benchmarking Open-Ended Multi-Agent Coordination in Language Agents — Open-ended multi-agent coordination is a distinct bottleneck for frontier LLMs, measurable by the $alem$ benchmark.
• not much happened today — Agentic AI development is converging on integrated stacks, robust infrastructure, and careful token economics for performance and cost.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.