Microsoft releases Agent Governance Toolkit

Why it matters

Microsoft has moved from talking about AI agent governance to shipping code for it. The company said on April 2 that it had released the Agent Governance Toolkit as an open-source project under the Microsoft organization on GitHub, describing it as runtime security governance for autonomous AI agents. (opensource.microsoft.com) That matters because Microsoft is framing agents as actors that need controls at execution time, not just guardrails in prompts or model settings. (github.com) The repository says the toolkit sits between an agent framework and the actions an agent takes, evaluating tool calls, resource access and inter-agent messages against policy before execution. ### What did Microsoft actually release? (devblogs.microsoft.com) Microsoft’s GitHub repository describes the Agent Governance Toolkit as a package for “policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous AI agents.” The project is published under the Microsoft organization and, according to Microsoft’s open-source blog, under the MIT license. April 2 is the date Microsoft used for the formal announcement on its open-source blog. (opensource.microsoft.com) In that post, the company said the toolkit was intended to bring runtime security governance to agent systems as regulation and enterprise adoption accelerate. ### What problem is the toolkit supposed to solve? (github.com) Microsoft said the governance gap is that agents are easy to build but harder to control once they can call tools, access data and coordinate with other agents. Its documentation says every action can be checked against deterministic policy before execution, creating an enforcement layer separate from the model’s own behavior. (github.com) The company’s Azure adoption guidance makes the same point in broader terms. Microsoft says organizations need governance and security practices for AI agents across the organization, with policy formation treated as a core step in agent adoption. (opensource.microsoft.com) ### Which controls are in the toolkit? The GitHub materials list policy enforcement, zero-trust identity, tamper-proof audit logging, prompt injection detection, rate limiting, circuit breakers, execution rings and OpenTelemetry metrics among the toolkit’s features. A.NET package page also describes the package as combining those controls in a single SDK. GitHub examples show Microsoft has built integrations around common agent frameworks. (opensource.microsoft.com) One example covers a CrewAI deployment, and community materials in the repository reference LangChain and OpenAI Agents as starting points for adding governance to existing projects. ### How is Microsoft positioning this in the agent stack? Microsoft said in a May blog post on its Agent Framework site that Agent Framework and the Agent Governance Toolkit are meant to work together as a production stack. (learn.microsoft.com) The company described Agent Framework 1.0 as the layer for building and orchestrating multi-agent systems, while the governance toolkit handles deterministic policy enforcement, intent-based authorization and auditable lineage. (github.com) That positioning matches the repository’s architecture language. GitHub documentation says the toolkit is designed to govern tool use, resource access and agent-to-agent communication, rather than replace the application framework itself. (github.com) ### Is this a finished product or still early? Microsoft’s release notes say the packages are “public preview” releases, even though they are Microsoft-signed and described as production-quality. The v3.6.0 notes dated May 12 said the releases may still have breaking changes before general availability, and the latest listed release was v3.7.0 last week. GitHub activity also shows the project is still being expanded. Repository pages list recent updates across tutorials, integrations, compliance mappings and command-line governance packages, suggesting Microsoft is continuing to add examples and packaging around the core enforcement layer. (devblogs.microsoft.com) ### What comes next? Microsoft’s adoption site is already using the toolkit as part of a broader agent-governance push, including webinars and governance guidance published in May. (github.com) The next visible milestones are likely to come through the GitHub releases page, where Microsoft has been posting versioned updates through May 2026, and through additional Agent Framework integrations the company is documenting publicly. (adoption.microsoft.com) (github.com 1) (github.com 2)

Key numbers

• Microsoft published its open-source Agent Governance Toolkit on April 2, 2026, positioning it as runtime security infrastructure for autonomous AI agents.
• (opensource.microsoft.com) The GitHub repository says the toolkit covers all 10 categories in the OWASP Agentic Top 10 and adds policy enforcement, zero-trust identity and audit logging.
• (github.com) Microsoft’s latest public materials tie the toolkit to Agent Framework integrations and GitHub examples for CrewAI, with releases continuing in May 2026.
• The company said on April 2 that it had released the Agent Governance Toolkit as an open-source project under the Microsoft organization on GitHub, describing it as runtime security governance for autonomous AI agents.

What happens next

• Microsoft said in a May blog post on its Agent Framework site that Agent Framework and the Agent Governance Toolkit are meant to work together as a production stack.
• The v3.6.0 notes dated May 12 said the releases may still have breaking changes before general availability, and the latest listed release was v3.7.0 last week.
• (devblogs.microsoft.com) What comes next?

Sources

• opensource.microsoft.com
• github.com
• devblogs.microsoft.com
• learn.microsoft.com
• github.com
• github.com
• adoption.microsoft.com
• github.com 1
• github.com 2

Quick answers