Chapter 1 of 6
Module scope#
What we will do#
Objective. Add operating signals to one governed Foundry project.
Deploy a diagnostic setting, an action group, and three platform metric alerts for the Foundry account. Then create or update one continuous evaluation rule for the governed agent. The result is an enabled alert set, a recorded evaluator decision, and a rule that scores sampled agent response completions.
Why it matters#
Problem. Offline evaluation gates catch known release regressions, but they do not watch live traffic. Operators still need to see throttling, latency, token volume, and quality drift after the agent is in use.
Solution. This module adds those live controls. Azure Monitor routes platform alerts to the operating owner, and Foundry continuous evaluation scores sampled responses with the approved evaluators.
EU AI Act. For high-risk AI systems, this module supports Article 12 (record-keeping) with diagnostic logs, Article 26(5) (deployer monitoring) with platform alerts, and Article 72 (post-market monitoring) with continuous evaluation. This is an engineering mapping, not legal advice.
Boundaries#
This module sits outside the numbered session sequence.
It changes one Foundry account's diagnostic setting, one action group, three Azure Monitor metric alerts, and one continuous evaluation rule for the approved project and agent. Azure Monitor and Foundry hold the live state. This repository holds the Bicep definition, parameter contract, evaluation decision record, operational script, and KQL check.
Agent identity, runtime baseline, release quality gates, incident routing, budgets, and promotion remain with Sessions 04, 09, 11, and 12.
Module preparation
Who should join
- Microsoft Foundry platform and agent owners
- Azure Monitor, Application Insights, and Log Analytics operators
- AI quality owners who operate continuous evaluation
- Service owners who respond to throttling, latency, and token-use alerts
What you need
- Session 04 has produced one governed Foundry agent in an approved nonproduction project.
- Session 09 has selected the evaluator set and quality thresholds for the same agent.
- Session 11 has named the Log Analytics workspace, Application Insights resource, alert owner, and operating route.
- Session 12 or the approved deployment path can run Bicep what-if and deploy to the exact resource group that contains the Foundry account.
- The operator has Monitoring Contributor on the approved resource group and the Foundry project managed identity can receive the Foundry User role through the customer's access process.