Does giving LLM teams a designated leader improve coordination?

The paper builds an embodied multi-agent framework, extending Zhang et al.'s architecture via the AutoGen conversation library, in which three or more LLM agents cooperate inside a physical/simulated environment (VirtualHome), alternating between a standalone communication phase and an action phase, with separate Actor and Communicator LLMs per agent. It compares teams under different prompted organizational structures: a "disorganized" team with no imposed structure, and a hierarchical team with a designated or elected leader. Measured behavior (Figure 7) shows disorganized teams spend the majority of their communication telling others what to do — "Leadership & assistance accounts for around > 50% of all the behavior" — while rarely requesting guidance ("request for guidance in < 10% of the dialogues"). In hierarchical teams the leader dominates giving orders ("amount to > 60% of their communication") while followers "tend to follow and give fewer orders" and "share and ask for more information." The paper concludes that "a hierarchically-organized team with a designated/elected leader has superior team efficiency."

The mechanism is purely textual: organizational structure is imposed by "including a textual description as part of the prompts for both the Actor and Communicator," not by hard-coding who may message whom. Communication is also structurally separated from action decisions — a distinct phase, a distinct LLM per agent — and agents can broadcast, address one or several specific recipients, or stay silent. The paper frames the underlying problem this is meant to fix as a known LLM tendency: "LLM agents tend to over-report and comply with any instruction, which may result in information redundancy and confusion in multi-agent cooperation." A role prompt (leader) and a dedicated coordination phase are enough to reorganize that tendency into fewer, more purposeful messages, without changing what any agent is technically permitted to do.

This sits alongside Does structured artifact sharing outperform conversational coordination?, which found that replacing free natural-language messages with structured artifacts reduces coordination failure. This paper shows a second, independent lever on the same problem: even leaving communication as free natural language, simply prompting a role/structure (a leader) onto the team measurably changes behavior and reduces chaos. The disorganized team's pattern — agents telling others what to do while rarely asking for guidance — is also a concrete behavioral instance of the failure described in Why don't LLM agents naturally explore each other in teams?, where peer interaction is myopic and commitment is premature unless agents are explicitly structured to seek information from each other.

The excerpt does not establish that this benefit generalizes: the experiments run in a single environment (VirtualHome), the paper states explicitly that the "current work... lacks human evaluation," and the environment "cannot hold hundreds of agents," so scaling to larger or more realistic organizations is untested. The finding supports "a leader-role prompt improves coordination" as true for small embodied teams in this one testbed and this family of models, not yet a general principle that role-based organizational prompting outperforms disorganization at scale or across environments.