Could AI teammates work better by trading structured documents instead of chatting back and forth in plain language?

This explores whether agents should coordinate by handing each other structured work products (specs, documents, schemas) instead of talking things through in natural language. The strongest evidence says yes, at least for well-understood workflows. MetaGPT gives agents the roles of a software team and has each one produce standardized engineering documents, the way a product manager writes a requirements doc and an architect writes a design spec. Agents then pull what they need from a shared workspace, and this beats agents chatting with each other Does structured artifact sharing outperform conversational coordination?. The reason is less about format than about noise. Conversation piles up irrelevant back-and-forth, while artifacts carry only what the next step needs. The same idea shows up inside a single agent. DeepAgent compresses its own messy history into structured memory schemas, and that both saves tokens and gives the agent a clean point to stop and rethink its strategy Can agents compress their own memory without losing critical details?.

Structure doesn't fix everything, though. The AgentsNet benchmark finds that coordination breaks down as networks grow for two reasons. Agents agree on a plan too late, or they adopt a strategy without telling their neighbors. They also accept what other agents tell them without checking it Why do multi-agent systems fail to coordinate at scale?. A well-formatted artifact can still be wrong, and a shared document can spread an error faster than a chat would. Structure solves the noise problem but not the trust problem. That may be why some researchers argue coordination needs outside infrastructure that tracks which agent did what and catches harm when it happens What infrastructure must exist beyond training agents?. Structured artifacts make that tracking much easier, since a document has an author and a version in a way a stream of messages doesn't.

A more interesting point is that natural language is being squeezed from two directions at once. On one side is more structure. The coordination pattern itself can be treated as an artifact, a graph whose nodes are prompts and whose edges define who passes what to whom, so both can be tuned automatically Can we automatically optimize both prompts and agent coordination?. Even finding the right agent for a task can be handled by structured capability descriptions that are matched automatically, so nobody has to describe or hand-wire every connection Can semantic capability vectors replace manual agent routing?. On the other side is less structure. Some research has agents skip words entirely and share latent "thoughts" pulled straight from their hidden internal states, which can expose disagreements before they ever show up in text Can agents share thoughts directly without using language?. In both cases, natural language starts to look like a convenient middle layer more than the thing coordination actually depends on.

There is a sobering caveat. One analysis finds that about 80% of the variation in multi-agent performance comes from how many tokens the system spends, not from how cleverly it coordinates How does test-time scaling work at the agent level?. Part of what artifacts buy you may simply be spending tokens more efficiently. The history of protocols also suggests that "replace" is the wrong frame. Coordination standards catch on by wrapping and connecting existing systems, not by displacing them Should coordination protocols wrap existing systems or replace them?. The likely future is layered: structured artifacts for handoffs and records, latent or graph-level channels for efficiency, and natural language kept for the open-ended parts no schema anticipated.