Trust-Aware Capability Discovery: Building Reliable Multi-Agent Systems on Azure

This session presents a trust-aware framework, grounded in real-world Azure AI Agent and Copilot Studio implementations, that complements existing agent communication protocols by introducing verifiable capability descriptions, challenge-based validation, reputation tracking, and mechanisms to detect changes in agent behavior. Rather than relying solely on self-declared capabilities, this approach surfaces measurable trust signals that help orchestrators make better-informed delegation decisions, while remaining fully compatible with existing MCP and A2A implementations already in use across Azure agentic workflows. The talk also explores how trust can be established across heterogeneous agent ecosystems spanning different models, architectures, and performance profiles, a common reality in enterprise Azure environments that mix Azure OpenAI-backed agents, Copilot Apps, and third-party agentic tools. We'll cover practical considerations for capability verification, delegation reliability, and reputation management, and discuss how protocol-level trust supports dependable agent collaboration at enterprise scale. The framework is designed as an incremental layer that plugs into existing Azure agentic deployments without requiring changes to underlying models or communication standards.