microsoft/agent-governance-toolkit

AGT solves the fundamental problem that prompt-level safety is insufficient for autonomous agents, as prompt injection attacks achieve near-100% success rates. It replaces probabilistic model-level controls with deterministic, code-enforced policies that make unauthorized actions structurally impossible, providing a verifiable security boundary.

Developers can build multi-agent systems with fine-grained access control, zero-trust identity, and tamper-evident audit logs. Real-world use cases include enterprise automation where agents interact with databases, email, and APIs, requiring strict policy enforcement. The framework supports any agent framework and enables compliance with regulatory requirements for auditability and incident response.