Microsoft Launches Open-Source AGT With 13,000 Tests to Govern AI Agents

6 articles · Updated · InfoWorld · May 28

Public-preview AGT wraps policy checks around agent actions before execution, letting developers allow, alert on, or block operations and log each decision.
Microsoft pitched the toolkit as a response to agentic AI driving massive API traffic, token overspending and risks such as goal hijacking, insecure output handling and uncontrolled code execution.
Less than 0.1ms policy evaluation per operation is meant to keep overhead low while AGT adds throttling, token budgets, drift detection, kill switches and audit trails.
Vendor-neutral support spans Azure Foundry, Amazon Bedrock, Google ADK, 19 framework adapters and five languages, with Python offering the fullest implementation.
The release signals a push to treat agents as a distinct class of software user as enterprises and regulators demand tighter cost, security and compliance controls.

Microsoft’s toolkit promises to control AI agent costs, but what is the hidden price of this new enterprise-grade governance?

AGT tackles top AI security risks, but as it's not a 'true security boundary', are businesses adopting a false sense of security?