AI Agents Are Forming Their Own Coalitions — No Central Commander Needed
Decentralized coalition formation using local payoff evaluation—no central planner needed.
This paper introduces a decentralized coalition-formation model where agents unilaterally exit or join groups, evaluating moves using the Aumann-Dreze value within their current coalition. A terminal partition occurs when no individually profitable exit-and-join deviation exists. The study establishes equilibrium characterizations, identifies conditions for scalar Lyapunov or exact-potential representations, and analyzes how switching and acceptance costs affect local stability. Numerical experiments test finite-time stabilization, cost sensitivity, and a convex-game benchmark.
- Agents use the Aumann-Dreze value for local payoff evaluation, not global negotiation.
- A terminal partition is reached when no unilateral exit-and-join move improves any agent's payoff.
- Cost sensitivity analysis shows switching and acceptance costs critically affect local stability and convergence.
Why It Matters
Enables scalable, fault-tolerant coalition formation for multi-agent AI systems without central coordination.