agent-governance-toolkit: Govern Autonomous AI Agent Actions

Overview

Agent Governance Toolkit (AGT) is a governance layer for autonomous AI agents. It intercepts actions such as tool calls and delegation so teams can evaluate them against policies, record decisions, and deny disallowed operations in application code rather than relying only on model instructions.

The toolkit is intended for teams building or operating agents across multiple frameworks. Its components cover policy enforcement, identity and trust, audit, execution controls, and reliability, with integrations and SDKs for several languages. It is not an operating-system security boundary.

Key Features

• Evaluate tool actions against YAML policies, including allow, deny, and approval outcomes.
• Record policy decisions in audit trails, with tamper-evident audit capabilities described by the project.
• Provide agent identity and trust features, including SPIFFE, DID, and mTLS support.
• Add execution controls such as privilege rings, sandboxing support, and termination controls.
• Integrate with agent frameworks including Microsoft Agent Framework, Semantic Kernel, LangGraph, CrewAI, and the OpenAI Agents SDK.
• Offer compliance checks and policy-linting tools through the agt CLI.
• Provide SDKs and packages for Python, TypeScript, .NET, Rust, and Go.

Use Cases

• An AI platform team wants to prevent agents from running destructive database operations or sending messages without required approval.
• A security team operating a multi-agent workflow needs to associate actions with agent identities and review decision records after an incident.
• A team using MCP servers wants to add governance checks and inspect for tool poisoning or hidden instructions.
• A compliance or risk team needs to map agent controls to frameworks such as OWASP Agentic AI Top 10, NIST AI RMF, or SOC 2.

Getting Started

Install the full Python package set:

pip install "agent-governance-toolkit[full]"

See the official repository and its README for setup details, framework integrations, and examples.

Considerations

• The project is labeled Public Preview, and breaking changes may occur before general availability.
• AGT enforces governance at the application middleware layer, not the OS kernel level. The agent and policy engine share a process boundary; the project recommends running each agent in a separate container for OS-level isolation.
• The README describes a broad set of packages and capabilities. Check the package matrix and documentation to confirm that a specific feature is available in the SDK or language you plan to use.
• The quick start requires Python 3.11 or newer, while the prerequisites section lists Python 3.10 or newer. Confirm the requirement for the distribution and release you install.
• Governance controls add policy configuration and operational work. They complement, rather than eliminate, the need for layered security and review of third-party services used by agents.