LLM agents should not talk to each other in English

A paper from Nova Berg argues that natural-language messages between LLM agents are an anti-pattern. The better approach is a shared structured clipboard that agents read from and write to. The problem is token waste. If one agent produces JSON and another agent needs to parse it, wrapping that JSON in conversational prose burns tokens on both sides. The producer agent uses tokens to format a polite message. The consumer agent uses tokens to parse that message back into structure. The clipboard pattern cuts that overhead. Agents write typed data to named slots. Other agents read from those slots. No “here is the data you requested” preamble. No “thank you for the information” acknowledgment. Just schema-validated writes and reads. This maps to how Unix pipes work. grep does not thank cat for the input. It reads stdin, does the work, writes stdout. The shell handles the plumbing. The paper shows latency drops and token costs drop when you strip the conversational layer. Agents still have natural-language interfaces for humans, but agent-to-agent communication is pure data. I have been guilty of building multi-agent systems that talk like humans because it felt intuitive. The clipboard pattern feels less intuitive until you remember that most successful distributed systems do not speak English to each other. They speak protocol buffers, JSON-RPC, or binary formats. LLM agents should do the same. The interesting question is whether this applies to chains that need context from previous steps. If the next agent needs to know why the previous agent made a decision, structured logs in the clipboard might be enough. If not, you are back to passing explanations, which brings the token overhead back.

---

Source: Natural-language messages between LLM agents are an architectural anti-pattern