Emergent Leadership and Frame Dynamics in Multi-Agent LLM Systems (Claude Opus 4.6)

I ran 1,000+ simulated social games using Claude Opus 4.6

and found that AI agent collectives develop structured leadership dynamics.

Core findings:

1. Influence creates a protective leadership shield

2. Consistency violation (not factual error) collapses leadership — 95% vs 10%

3. Leadership frames survive individual leaders through successor agents

4. Frame legitimacy depends on speaker type, not content alone

5. Triadic governance structures provide the highest collective stability (87.5%)

We propose the AI Leadership Equations — a formal model of collective stability.