Trust & Governance

Why governance is non-negotiable: In regulated industries (financial services, healthcare, government), deploying AI agents without auditable governance is a compliance violation waiting to happen. The Enterprise Kit makes governance the default, not an afterthought — every agent decision is tracked, every policy is enforced, and every promotion is gated through automated certification.

Overview

The Enterprise Kit embeds governance at every layer — from declarative policy packs that constrain agent behavior to an immutable decision ledger that records every action. Agent 365 (GA May 2026) serves as the enterprise governance control plane for all AI agents; this kit enhances Agent 365 with enterprise-grade controls that Microsoft does not provide natively. The Trust & Certification Plane provides the assurance framework for enterprise-grade AI deployments.

Trust Extensions

Decision Ledger

An immutable, hash-chained audit trail that records every significant agent decision. Each entry includes the agent identity, action taken, reasoning trace, policy evaluation results, and a SHA-256 hash linking to the previous record.

Tamper-evident — hash chain breaks reveal unauthorized modifications
Cosmos DB-backed with configurable retention policies
Query interface via LedgerReader for audit and compliance
JSON Schema-validated record format

Scorecard Compiler

Aggregates multiple quality signals into a single trust score per agent:

Evaluation results (accuracy, groundedness, relevance)
Policy compliance rate
Telemetry health (error rate, latency P99)
Human feedback scores
Weighted formula with configurable thresholds

Certification Pipeline

Automated agent promotion across environments with gate checks:

Tier	Environment	Requirements
Bronze	Staging	Manifest valid, all evaluators pass (>0.7), policy baseline compliance
Silver	Pre-production	Trust score >0.8, all policy packs pass strict mode, ledger chain intact, human review approval
Gold	Production	Trust score >0.9, all policies pass, 99th-percentile performance, external audit pass, full audit trail complete

Policy Graph

Visual compliance topology that maps relationships between agents, tools, and policies. Enables real-time violation detection and compliance reporting across the entire agent fleet.

Guardrails Manager (Azure Foundry Guardrails)

Integrates with Azure AI Foundry Guardrails (GA) — the platform-native content safety layer providing PII detection, prompt attack defense, task adherence monitoring, and content filtering at four intervention points (input, output, tool call, tool response). Supports configurable severity thresholds and custom blocklists via Azure AI Content Safety.

Agent Application Publisher

Deploys agents as first-class Azure Foundry Agent Applications (ARM resources) with full lifecycle management — create, update, version, and decommission.

Agent 365 Publisher

Registers governed agents in Microsoft Agent 365 — the cross-platform governance hub providing unified agent inventory, Teams/M365 Copilot distribution, lifecycle management, and compliance monitoring. The publisher syncs trust scores, certification levels, and policy compliance status with Agent 365 for enterprise-wide visibility.

MCP Fleet Governance

Enterprise-grade lifecycle management for MCP servers deployed at scale:

Toolbox Lifecycle Manager — Version, promote, deprecate, and retire tools across environments
MCP Auth Provider — Unified authentication for all MCP servers (Managed Identity, Entra Agent Identity, OAuth passthrough)
Private Tool Catalog — Azure API Center-backed discovery and approval workflow for MCP tools
MCP Hosting Templates — Azure Functions and Container Apps templates for hosting custom MCP servers
Fleet Dashboard — KQL-powered monitoring of tool call volumes, latencies, errors, and auth failures across all MCP servers

Policy Packs

Declarative policies defined in JSON/YAML that govern agent behavior:

Policy Pack	Purpose
Data Boundary	Geographic and classification constraints on data access
Identity (Entra ID)	RBAC roles, conditional access, MFA requirements
Escalation	Rules for when agents must escalate to humans
Memory Retention	TTL and purge rules for agent memory stores
Toolbox Allowlist	Per-agent tool access restrictions

Azure Governance Integration

Capability	Azure Service	Kit Component
Fleet Monitoring	Azure AI Foundry Control Plane	Policy Graph + Observability Suite
Agent 365 Governance	Microsoft Agent 365	Agent365 Publisher + trust score sync
MCP Fleet Governance	Foundry Toolbox + API Center	Toolbox Lifecycle Manager + Private Tool Catalog
Threat Detection	Microsoft Defender for AI Agents	Guardrails Manager alerts
Data Loss Prevention	Azure Purview	Data Boundary Policy integration
Audit Storage	Azure Cosmos DB	Decision Ledger (hash-chained)
Secret Signing	Azure Key Vault (RSA)	Ledger export verification
Identity & Access	Microsoft Entra ID	RBAC roles + Conditional Access
Content Safety	Azure AI Content Safety	Foundry Guardrails (4 intervention points)
Compliance Enforcement	Foundry Control Plane	Audit + Enforce modes

Compliance Workflow

Define policy packs in policy-packs/
Reference policies in agent manifests
Policy engine evaluates at deployment and runtime
Violations logged to Decision Ledger (Azure Cosmos DB)
Scorecard aggregates compliance metrics
Foundry Control Plane enforces fleet-wide compliance
Certification pipeline gates based on compliance score
Microsoft Defender alerts on anomalous agent behavior