The experiment evaluated 8 models across 4-256 agents and 8 coordination protocols ranging from externally imposed hierarchy to emergent self-organization (arXiv:2603.28990). Agents given only fixed ordering spontaneously developed specialized roles, voluntarily abstained from out-of-competence tasks, and formed shallow hierarchies without any pre-assigned roles or external design. Stronger models scaled emergent autonomy effectively while those below a capability threshold performed better under rigid structures.

The Sequential hybrid protocol outperformed centralized coordination by 14% (p<0.001), producing a 44% quality spread between protocols (Cohen's d=1.86, p<0.0001). The system scaled sub-linearly to 256 agents without quality degradation (p=0.61) and generated 5,006 unique roles from only 8 agents. Results replicated across closed- and open-source models, with the latter reaching 95% of closed-source quality at 24x lower cost.

The authors conclude that autonomous behavior already emerges in current LLM agents and that as foundation models improve, the scope for autonomous coordination will expand (arXiv:2603.28990). They state the practical implication is to provide agents a mission, a protocol, and a capable model rather than pre-assigned roles.