- A new arXiv paper analyzes how language-model agents talk to each other and finds free-form messages inflate token usage and cost.
- The authors propose PACT (Protocolized Action-state Communication and Transmission), which compresses each agent output into a compact action-state record.
- PACT lifts OpenHands’ resolve rate at 10% fewer tokens per resolved task and halves input tokens on SWE-agent while staying resolve-neutral.
- No single communication strategy was universally optimal across the topologies tested.
What Happened
Researchers Chen Huang, Yuhao Wu, and Wenxuan Zhang posted a paper, “What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems” (arXiv:2606.05304), submitted June 3, 2026. It targets a quiet cost driver in multi-agent systems: the unconstrained natural language agents pass to one another.
The abstract states that free-form communication “can rapidly inflate token usage, consume the shared context window, and ultimately affect both system performance and inference cost.”
Why It Matters
Multi-agent systems are increasingly the architecture behind autonomous coding and research tools, including the long-horizon agents measured by work like the DeepSWE benchmark. Token cost is the hidden tax on these systems, and shrinking it directly improves the economics of agentic AI.
Technical Details
The authors analyzed five common inter-agent communication strategies across two MAS topologies and found no fixed strategy was universally optimal; effective messages consistently preserved action-centered information. PACT treats communication as a public state-update problem, projecting each raw output into a compact action-state record before it enters shared history.
In production coding harnesses, PACT lifted OpenHands’ resolve rate at 10% fewer tokens per resolved task, and was resolve-neutral on SWE-agent while halving input tokens. The code is publicly available.
Who’s Affected
Developers building multi-agent coding and research systems are the direct audience, especially teams running high call volumes where input tokens dominate cost. The findings complement other agent-design research released the same week, including SentinelBench.
What’s Next
The results come from the authors’ own evaluation on two topologies and two coding harnesses; broader adoption depends on whether PACT’s gains generalize to other frameworks. The public code release lowers the barrier for independent replication.
