Capgemini Executives: Multi-Agent AI Governance Needs Three Control Layers, Not Explainability
Capgemini's Deputy Group CTO Sudhir Pai and AI Futures Lab head Mark Roberts argue that scaling multi-agent systems requires a three-level control architecture—per-agent harness, per-system control plane, and enterprise-wide control plane—and a shift from explainability to containment, because emergent interactions can produce wrong outcomes even when every individual agent behaves correctly.
Capgemini's Deputy Group CTO Sudhir Pai identifies a specific failure mode that distinguishes multi-agent production from single-agent testing: every agent can be right while the system is wrong. The problem is not behavioral drift in any one agent, but interactions nobody designed. As enterprises deploy growing numbers of agents, the systems were not built to test or govern emergent coordination.
Pai's proposed response is a three-level architecture. The first level is a harness around each individual agent. The second is a control plane for each agentic system. The third, and the one he argues will dominate 2027 enterprise agendas, is an enterprise control plane across all systems. This top layer must incorporate a register of every agent and agentic system, identity and permissions for each one, monitoring of outcomes delivered and consumption, visibility into how agents interact, limits on what they can do together, and a way to stop them.
Mark Roberts, Head of Capgemini's AI Futures Lab, frames the shift in safety philosophy. The traditional approach—limiting agent access and checking that every decision remains within business rules—assumes you can verify decisions by examining reasoning steps. But some new models no longer show all their reasoning steps. Roberts argues that even if you can explain each component, you still do not know how the overall system will behave. The question is not whether more explainability leads to more safety; it is whether containment at the macro level can substitute for understanding at the micro level. He draws an analogy to traffic systems: you do not need human oversight in every loop, but you need it where it matters, with layers of safeguards so mistakes are caught when they are small and recoverable.