Agent runtime on Azure AI Foundry
The MCP-based agent platform runs on Azure AI Foundry Agent Service, giving each agent a managed, isolated runtime instead of open-ended execution.
AI Platform | Microsoft Azure
AI agents needed real tools. The security team needed to sleep at night.
A governed MCP platform on Azure gave enterprise AI agents controlled, auditable access to production tools, and to Teams and Copilot.
Project brief
AI agents got only the tools they need, through one gateway, with every call logged.
A governed MCP platform on Azure gave enterprise AI agents controlled, auditable access to production tools, and to Teams and Copilot.
01 | Challenge
The operating problem.
The organization wanted AI agents and copilots to take real actions in production systems, such as querying data, triggering workflows and calling internal APIs.
The blocker was governance: direct agent access created serious risk around auditability, authentication and blast-radius control.
Leadership needed agent capability without giving autonomous systems unrestricted access to production.
02 | Delivery approach
01
02
03
04
05
How the solution was delivered.
Identity and access with Entra ID
Every agent and tool call authenticates with a scoped Entra ID OAuth token, so agents only reach the tools and data they are explicitly authorized to use.
Governed gateway with APIM
Production tools are exposed only through Azure API Management (APIM), so every call is mediated, rate-limited, versioned and never hits a backend directly.
Auditability through APIM and Entra ID logs
Every tool invocation is captured in APIM request logs, tied back to the calling agent's Entra ID token, creating a queryable audit trail of which agent did what, when and with what result.
Business access through Teams and Copilot
The same governed agents are surfaced to end users inside Microsoft Teams and Copilot, so the controls apply whether a request comes through an API or a chat interface.
Architecture
How a request moves through the platform, end to end.
The architecture view summarizes the workstream sequence, control points and technical path used to move from challenge to operating outcome.
Metrics
Published indicators
Where public numeric metrics are not approved, indicators are shared as operating signals.
Least privilege
Access model
Entra ID
Identity
APIM
Gateway
03 | Outcome
Production impact with client confidentiality protected.
Published outcomes remain qualitative unless client-approved metrics are available for public use.
Tech stack
Platforms and controls used in the engagement.
Azure AI Foundry Agent Service
Entra ID
Azure API Management (APIM)
Microsoft Teams
Microsoft 365 Copilot
Model Context Protocol (MCP)
AI & Platform Engineering
This case study connects directly to iNET SYSTEMS AI & Platform Engineering delivery.
Explore service
Related case studies
Related delivery stories.
Free first call
Book a Consultation