From One Directive to Full-Organization Action: Organizational Mirroring for Multi-Agent LLM Systems

Current multi-agent frameworks adopt flat communication topologies, requiring designers to manually orchestrate each agent interaction. We present organizational mirroring, an approach that allows a single natural-language directive to mobilize an entire LLM agent organization—mirroring how a CEO's briefing cascades through a real company.

The approach rests on eight principles: hierarchical delegation, independent memory, layered compression, meta-department, skill-based composition, self-evolution, replaceable executors, and real-world workflow mapping. We implement these in a production system of 18 agents across four departments on the OpenClaw framework.

Controlled experiments on a 30-task suite (90 runs across three topologies) show: (1) 3.1–5.9x greater output volume; (2) quality scores of 18.3/20; (3) hierarchical coordination as primary quality driver (Cohen's d=1.409); (4) 85.7% communication link reduction; (5) zero cross-domain lexical intrusion; (6) 87.5% inter-rater agreement between independent evaluation layers.

This archive contains both the English version (paper-EN.pdf) and the Chinese version (paper-CN.pdf) of the paper.