Abstract
AI agents are moving from demos to production — but who governs what they do at runtime? The Agent Governance Toolkit (AGT) is an open-source, MIT-licensed framework from Microsoft that enforces deterministic policy before every tool call, message, and action an agent takes. It’s not prompt engineering — it’s runtime governance.
In this talk, I’ll walk through how AGT brings zero-trust identity (Ed25519 + post-quantum ML-DSA-65), policy-as-code (YAML, OPA/Rego, Cedar), tamper-evident Merkle audit chains, and a Kubernetes sidecar model to any AI agent — regardless of framework. With SDKs in Python, TypeScript, .NET, Rust, and Go, and native integrations for LangChain, CrewAI, AutoGen, Semantic Kernel, MCP, and A2A, AGT slots into your existing stack. I’ll demo the OpenClaw sidecar on Ubuntu/AKS, show how policy evaluation stays under 5ms, and
share how we mapped all 10 OWASP Agentic risks and compliance frameworks like EU AI Act, SOC 2, and NIST AI RMF.
Attendees will leave with a concrete understanding of how to govern autonomous agents in production — and how to contribute.
Speaker Bio
Imran Siddique
Imran Siddique is a Principal Group Engineering Manager and Agentic AI Architect at Microsoft. He is the creator of the Agent Governance Toolkit and the “Scale by Subtraction” philosophy for hyper-scale systems. A holder of multiple patents, Imran currently leads the engineering of backend services and Knowledge Graphs powering next-generation AI agents.