Governing at the Speed of Agents: Sandboxes, Anti-Capture Safeguards, and Institutional Capacity for Agentic AI in Healthcare

National healthcare AI guidelines are typically revised on multi-year cycles, yet the state of AI they regulate can change rapidly within months. We use Singapore's Artificial Intelligence in Healthcare Guidelines (AIHGle), updated from Version 1.0 (October 2021) to Version 2.0 (March 2026), as a case study of this cadence mismatch. We set it against two 2026 incidents in which frontier models from Anthropic and OpenAI autonomously breached the boundary between sandboxed evaluation environments and live production systems during routine safety testing. We argue that periodic, document-based guideline revision alone cannot keep pace with agentic risk, however frequently it is updated, and that the remedy is not simply faster rules but layered governance: a stable core of risk-based duties and outcomes paired with modular technical controls and standing infrastructure that can be revised on the technology's own timescale. In this context, we propose three interventions: (1) standing, co-created regulatory and implementation sandboxes that give regulators, developers, deployers, and patients shared, continuously updated evidence about agentic behaviour; (2) cautious assessment of new safety requirement, including calls for regulation from frontier developers themselves, so that safety measures do not inadvertently entrench incumbents; and (3) deliberate investment in institutional capacity so clinicians, deployers, and regulators can meaningfully interrogate agentic systems. Through these proposals, we aim to allow policy and regulation to keep pace with the agentic AI technological evolution. For submission under Topic 2 "Trustworthy AI, Policy, Human-AI Collaboration & Adoption" and Submission Track "Position Papers".