What happened
Microsoft published the open-source agent-governance-toolkit (public-preview, on GitHub and PyPI) as a framework-agnostic governance layer for autonomous agents. Its runtime policy engine intercepts every tool call, message send and delegation in deterministic application code (so denied actions are structurally impossible), adds zero-trust agent identity, execution sandboxing, tamper-evident audit logging, and maps to OWASP Agentic Top 10. Multi-language SDKs (Python, TypeScript, Go, Rust, .NET) and adapters for LangChain, CrewAI, OpenAI Agents SDK, LangGraph and Microsoft Agent Framework.
Why it matters
This is a from-Microsoft, framework-agnostic, deterministic fail-closed governance surface that treats prompt-level safety as insufficient — an important open-source reference implementation for agent policy enforcement, identity and audit that enterprises can adopt directly.
Applicability
Platform and security teams building custom agentic systems should evaluate AGT (currently public preview) as the enforcement layer for tool calls and delegation, and watch for GA and breaking-change hardening.