AINEWS
The Cost of Chatter: Why Multi-Agent Systems Need Tighter Message Protocols
An arXiv preprint argues the cheapest multi agent wins come from tightening the message protocol — not from upgrading the underlying model.
An arXiv preprint argues the cheapest multi agent wins come from tightening the message protocol — not from upgrading the underlying model.
A new arXiv preprint argues that the cheapest wins in multi-agent LLM systems are hiding in plain sight — in the messages agents pass to each other, not in the models doing the passing.
The paper, "What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems" (arXiv:2606.05304, v1 submitted 3 June 2026), is the work of Chen Huang, Yuhao Wu, and Wenxuan Zhang. Its central claim is uncomfortable for anyone who has been treating inter-agent communication as a free-form chat channel: most of the tokens those messages consume are not doing the work downstream agents actually need, and that overhead shows up both on the bill and in the performance of the system as a whole.
In a typical LLM-based multi-agent system, agents are organized by role, by pipeline, or by turn schedule, but the content they exchange is usually unconstrained natural language. That choice is convenient — agents can "say" whatever they want — but it has costs that compound at every step.
Three things happen when messages are free-form:
The paper frames this as a problem of message design rather than model design: before any agent gets any smarter, the channel it talks on is already spending tokens on noise.
To pin down where the waste is, the authors ran an ablation across five common inter-agent communication strategies on two multi-agent topologies. The result that survives that comparison is the one worth paying attention to: no single fixed strategy is universally optimal. Some strategies win on one topology, others on the other, and the right choice depends on the task.
But a more durable signal also emerged. Across the strategies and topologies the authors tested, the messages that actually helped downstream agents were the ones that preserved action-centered information — what the upstream agent did, what state it left behind, and what the next agent needs to act on. Messages that wandered into commentary, justification, or restatement of context did not pay for themselves.
That finding, more than any specific strategy, is the constructive claim of the paper: a useful inter-agent message is a record of action and state, not a paragraph of prose.
The paper's proposed artifact is PACT — Protocolized Action-state Communication and Transmission — a harness-level pattern that treats inter-agent communication as a public state-update problem. Instead of letting each agent emit a free-form message, PACT projects the raw output into a compact action-state record before that record enters the shared history.
The release is public: the iNLP-Lab/PACT repository on GitHub ships the implementation, the harness integration, and the configurations used in the reported experiments. (The "iNLP-Lab" affiliation is inferred from the repository URL; the arXiv abstract page does not state the authors' institution explicitly, so treat that label as tentative.)
The mechanism is straightforward. Each agent still reasons in natural language privately, but the message it hands to the next agent is a structured record — what action was taken, what the resulting state is, and what the next agent should treat as given. Downstream agents read records, not paragraphs.
The strongest reported numbers in the paper come from running PACT through two production coding harnesses. Both are self-reported by the authors on the harnesses and configurations they selected; no independent replication appears in the current source set, so these should be read as the authors' own measurements rather than a settled benchmark.
Both results point in the same direction. The lever is the message protocol, not the model.
Two honest caveats sit alongside the numbers.
First, the paper is an arXiv preprint (v1, submitted 3 June 2026) and has not been peer-reviewed; the arXiv abstract page should be cited as a preprint, not a publication.
Second, the paper itself is clear that PACT does not replace good agent design. Its empirical finding is that no fixed communication strategy is universally optimal — the gain comes from a pattern (action-centering, protocolized records) that fits the workload, not from a one-size-fits-all template. The repository and the paper frame PACT as a complement to careful agent design, not a substitute for it.
The broader point is the one Type0 engineers should care about. The paper's evidence suggests that the performance-cost frontier of a multi-agent system can be moved by changing the harness, without retraining the underlying model. For teams already running multi-agent workflows on OpenHands, SWE-agent, or comparable harnesses, action-centered message records are a testable pattern, and the code in the repository is the place to start.