Key takeaway: Structured, action-oriented agent communication (PACT) reduces token costs in multi-agent systems without performance loss.
Researchers have analyzed how agents in LLM-based systems should communicate with each other to reduce token consumption. With the PACT protocol, performance can be maintained or improved — with significantly lower token throughput.
In multi-agent systems based on large language models, communication between agents is typically realized as free natural language outputs. This leads to uncontrolled token consumption, strains the shared context window, and increases inference costs without corresponding performance gains.
The work analyzes five common inter-agent communication strategies across two different multi-agent topologies. The key finding: there is no universally optimal strategy. However, it consistently shows that successful messages between agents contain action-oriented information that downstream agents need. This insight leads to the PACT protocol (Protocolized Action-state Communication and Transmission).
PACT treats inter-agent communication as a public state-update problem. Each agent output is projected into a compressed action-state record before it enters the shared history. This creates a structured message focused on essential information instead of free text.
Measurements consistently show improved performance-cost trade-offs across different multi-agent topologies: comparable or stronger task performance with significantly fewer tokens. In practical code systems: OpenHands achieves higher resolve rate with 10% fewer tokens per resolved issue; SWE-Agent remains performance-neutral but requires only half the input tokens. The code is publicly available at https://github.com/iNLP-Lab/PACT.
Source: arxiv.org · Published June 2, 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrase and classification through Lumi News Pipeline v1.6.5.