# Agent Governance Toolkit

A Python toolkit that intercepts every AI agent tool call, message, and delegation in application code to enforce policy, identity, and sandboxing before it happens.

The Agent Governance Toolkit (AGT) answers three questions about an autonomous agent in production: whether a specific action is allowed regardless of what a broader API scope or IAM role permits, which specific agent performed an action when several share credentials, and whether a tamper-evident record exists proving what policy was active and why a request was allowed or denied. Rather than relying on the model to follow instructions in its prompt, which published research reports can be defeated with a high success rate through adaptive jailbreak attacks, AGT intercepts every tool call, message send, and delegation in deterministic application code before the model's intent reaches the network.

It targets the OWASP Agentic Top 10 risks and installs as a single pip package that works with any agent framework. The project describes itself as a production-quality public preview that may still include breaking changes before a general-availability release.

It is written in Python and distributed under the MIT license.

Website: https://github.com/microsoft/agent-governance-toolkit