Agentic AI Governance
6 steps1 pieces6 Sources
The core challenge of agentic AI governance is shifting from reviewing model outputs to controlling autonomous actions. As multi-agent systems scale, emergent interactions can produce harmful outcomes even when individual agents behave correctly, creating a "governability gap" where capability outpaces control. Recent frameworks propose layered architectures—such as Capgemini’s three-level control plane and Futurum’s five-layer reference model—to enforce authority envelopes, traceability, and containment at runtime. Regulatory bodies like Singapore’s IMDA and the HKMA are issuing voluntary guidance and discussion papers emphasizing human accountability and risk-integrated harnesses. The disagreement lies in whether explainability or structural containment is the primary safety mechanism; consensus is emerging that macro-level control planes are necessary to manage agent-to-agent propagation and delegated authority.
How it got here
2025-08-05: Charles L. Wang et al. submit the MI9 paper on arXiv, proposing an integrated runtime governance framework with six components including agency-risk index and graduated containment strategies.
arxiv.org ↗2026-04-03: Futurum Group releases the Agent Control Plane Framework (ACPF), defining a five-layer architecture to bridge the gap between agent capability and governability for production deployment.
futurumgroup.com ↗2026-05-20: Singapore’s IMDA publishes Version 1.5 of the Model AI Governance Framework for Agentic AI, offering voluntary guidance on four dimensions including risk bounding and human accountability.