Deploy-Ready Agent Governance: App Service Integration and the Architecture That Actually Scales

Analysis

Where the Agent Governance Toolkit Falls Short: Gaps a Skeptic Should Know

The Awesome Agents writeup is the most balanced take on AGT so far, flagging that the semantic intent classifier has no independent red-team validation, sample configs are not production-ready, and the toolkit is still public preview with no stable v1.0 timeline. Worth reading alongside the official Microsoft posts before you make a production commitment.

News

OWASP ASI 2026 Taxonomy Now Drives AGT's Risk Coverage

The updated architecture post confirms AGT now uses the ASI01-ASI10 identifiers from the OWASP Agentic Security Initiative's 2026 taxonomy, with backward compatibility for the original AT numbering. This matters because compliance teams mapping to OWASP will need to track which identifier set their tooling and auditors are using.

Tools

Cross-Model Verification Kernel: How AGT Handles Memory Poisoning

The Help Net Security breakdown surfaces a detail that most coverage skips: AGT's memory poisoning defense uses a Cross-Model Verification Kernel with majority voting across models, not just a single classifier. That architectural choice has real implications for latency and cost that builders should factor in before enabling it.