Federation Is the Way Forward for AI Agents
Herbert Woisetschläger, Nicholas D. Lane, Shiqiang Wang
Abstract
Current AI agents are largely deployed as centralized services. This architecture is prone to correlated reliability failures, unsustainable energy concentration, and structural barriers to accessing high-value private context distributed across institutions and jurisdictions. In this position paper, we argue that federation is the way forward for AI agents, as the most valuable context for agents is often scattered across many entities. Hence, we believe that federated coordination should become the control plane of agentic systems, where using a high-capability centralized model is one of its possible decisions. We support our position through technical, policy, and economic analysis, showing that the shift towards federated agentic architectures is both necessary and tractable. We project that federation leads to significant GDP growth potential of $402B and $304B over 10 years for the U.S. and EU, respectively. These effects are enabled by unlocking cross-boundary tasks that centralized architectures cannot reach. We also propose a deployment roadmap and research agenda that directly address the most significant federation risks.