Organizations are increasingly deploying autonomous AI agents with broad access to production systems and sensitive data. These agents, while efficient, pose significant security risks due to potential vulnerabilities like prompt injection and LLM non-determinism. Relying solely on system prompts for security is insufficient, as they offer only soft boundaries and lack deterministic enforcement. To address this, Microsoft introduced the Agent Governance Toolkit (AGT), an open-source solution for production-grade agent governance.AGT acts as an interception layer between the AI agent's decision-making and actual tool execution. It enforces deterministic policies, preventing unauthorized or destructive operations before they occur. Unlike traditional security measures that fail to inspect downstream intent, AGT examines and controls agent actions. This approach is crucial given the evolving threat landscape, including OWASP Top 10 for Agentic AI vulnerabilities and regulatory mandates like the EU AI Act.The Agent Governance Toolkit comprises several components, including a Policy Enforcement Engine, a Cryptographic Identity layer, Runtime Isolation, and Tamper-Evident Logging. It operates as a modular control plane, decoupling policy enforcement from agent frameworks. AGT supports multiple SDKs and integrates with popular agentic frameworks, allowing for the retrofitting of existing systems.A hands-on demonstration shows AGT's real-time policy enforcement in Python. By defining rules in simple YAML, developers can wrap tool functions to automatically intercept and deny prohibited actions. This ensures that operations like "delete" are structurally impossible to execute without proper authorization, independent of the LLM's logic. AGT also provides a verification command to generate compliance reports and badges, offering evidence of security posture for review.