Can LLMs automatically discover better team structures than humans?

The paper introduces "Criticize-Reflect," a dual-LLM process that iteratively rewrites the team's organizational prompt rather than the task prompt. An LLM critic evaluates one episode's dialogue and action history, "analyzes the input and reasons to extract and summarize the key steps that are believed to influence the performance," and ranks agents' leadership. An LLM coordinator then takes the critic's output together with cost metrics — "time to task completion and communication cost" — from prior episodes, reflects on them, and "proposes a new and different organizational prompt for the next episode." Run across episodes, the loop "discovers more effective, novel organizational structures with self-improvement," and the paper reports three such structures that beat hand-designed alternatives in their respective settings: "(c) chain, (d) dual-leader, and (e) dynamic structures." The paper adds that the process can generate prompts implying "novel organizational structures that are rarely seen in human societies."

The mechanism is an automatic prompt-optimization loop aimed specifically at the organizational/role-structure text, using verbal, natural-language feedback rather than gradient-based optimization — explicitly modeled on the Actor-Critic pattern from reinforcement learning, with the critic supplying qualitative evaluation and the coordinator supplying the next candidate structure. Because the object being optimized is the structure prompt itself (who leads, who follows, how communication is patterned) rather than any single agent's task instructions, the search operates one level up from ordinary prompt engineering.

This is a concrete instance of the "edge optimization" described in Can we automatically optimize both prompts and agent coordination?: Criticize-Reflect optimizes who talks to whom via LLM-generated critique rather than graph search over a computational graph. It also parallels Do LLM agents develop unreadable languages when communicating under pressure? — another case where optimization pressure on LLM agent communication produces structures that diverge from human norms and are not directly human-legible, though here the pressure is explicit iterative critique rather than emergent task pressure, and the output is a team structure rather than a language. The structures this loop discovers build directly on Does giving LLM teams a designated leader improve coordination?, iterating past that paper's disorganized-versus-hierarchical baseline toward chain, dual-leader, and dynamic variants.

The excerpt does not establish that the discovered structures generalize beyond the single VirtualHome testbed used here; the paper states the "current work is performed in a single environment and lacks human evaluation." Nor does it establish that "rarely seen in human societies" means the structures are superior for any reason beyond the two measured proxies (time to completion and communication cost) in this one setting — only that an automated critique loop can out-search manual organizational-prompt design on those two metrics here.