Microsoft has released the agent-governance-toolkit, a public preview framework designed to provide deterministic policy enforcement, identity management, and auditing for autonomous AI agents. Moving beyond probabilistic prompt-level safety—which research shows can result in a 100% attack success rate against models like GPT-4o and Llama-3—the toolkit intercepts tool calls, messages, and delegations in application code to ensure actions are structurally impossible if they violate predefined rules. The toolkit utilizes a "fail-closed" architecture where the policy engine evaluates YAML-based rules to allow, deny, or require human approval for specific actions, such as blocking destructive database operations. It supports multiple programming languages, including Python (via the agent-governance-toolkit[full] package), TypeScript, .NET, Rust, and Go, and integrates with major frameworks like AutoGen, LangChain, and Semantic Kernel. Key components include the Agent OS for lifecycle management, the Agent Mesh for identity and trust, and an Agent Hypervisor for execution control. The toolkit is designed to align with regulatory and security standards, including OWASP LLM01:2025, NIST AI RMF 1.0, and the EU AI Act, providing tamper-evident audit logs for compliance and incident response.