Old theories explain how human teams coordinate — do they also explain how teams of AI agents should work together?

This explores whether decades of thinking about how teams and organizations coordinate still applies to today's LLM agent teams. The short answer from the corpus: yes, and the systems that borrow from it most directly tend to work best. The more surprising finding is that many current multi-agent setups aren't coordinating much at all.

Start with the borrowing. Organizational theory from the 1990s and 2000s already had clean answers to problems LLM builders keep running into. MOISE+, a 2002 model, splits an organization into three separate pieces: who holds which roles, what the overall plan is, and what each role is allowed or required to do. Keeping these apart stops agents from ending up with no guidance, or from losing a plan everyone already agreed on Can separating structure and function improve multiagent organization design?. IMACS uses the same idea in an LLM setting. It pulls team roles from Belbin's management research and coordination styles from Mintzberg's organizational theory, then treats them as swappable layers, so you can change the team's structure without rewriting how it coordinates Can multi-agent teams swap roles, coordination, and algorithms independently?. MetaGPT goes further and copies human workplace habits directly. Agents hand each other standardized documents, much like the paperwork in an engineering team, rather than chatting, and they coordinate better that way Does structured artifact sharing outperform conversational coordination?.

Now the problem. Coordination theory assumes the participants are really coordinating. Several findings suggest LLM teams often aren't. Token usage explains about 80% of the performance differences between multi-agent systems. These systems use around 15× more tokens than a single agent, and above a certain accuracy level, adding coordination makes results worse Are multi-agent systems actually intelligent coordination or just token spending? How does test-time scaling work at the agent level?. Put simply, much of what looks like teamwork is parallel spending. When researchers test coordination directly, agents fail in ways a coordination theorist would recognize right away. They agree too late, adopt a strategy without telling their neighbors, and accept information from peers without checking it, so errors spread through the network as it grows Why do multi-agent systems fail to coordinate at scale?. Even "socializing" is shallow: agents change their behavior when they know peers are present, but their language and ideas don't converge Do AI agents actually socialize with each other?.

That gap is why the theory matters most at the level of infrastructure, not individual agents. One argument holds that single agents hit an organizational ceiling that more capability can't fix. Tasks that need mixed expertise, parallel work, and independent checking require an explicit structure that spreads work across agents Do single agents always hit organizational limits?. Others argue that structure has to live outside the agents. That means systems that record which agent did what, shape how agents interact, and catch and fix harm, modeled on internet protocols What infrastructure must exist beyond training agents?. It also means supervisors that can halt an agent from outside its loop, because instructions in the prompt alone can't guarantee an agent stops Can prompt alignment alone guarantee agent termination in loops?. Standards that wrap existing protocols like MCP, rather than replacing them, are the ones that get adopted Should coordination protocols wrap existing systems or replace them?.

The takeaway you might not expect: coordination theory may be more useful to LLM agents than to the human organizations it was written for. People coordinate informally whether or not anyone designs it. LLM agents don't, so roles, obligations, shared documents, and outside oversight have to be built in explicitly. Otherwise a "multi-agent system" is mostly an expensive way to run the same model many times. One promising direction skips natural language between agents entirely and has them share internal states instead What makes multi-agent teams actually perform better?.