AI
Evaluating Agents Before They Ship
The Azure AI Evaluation SDK covers three categories: quality (fluency, relevance, groundedness), safety (violence, self-harm, sexual content, indirect jailbreaks), and a new agentic tier. Agentic evaluators check tool call accuracy, intent resolution, and task adherence. These dimensions matter because agents call external tools autonomously, and a standard quality score won’t catch a tool being called with wrong parameters. Custom evaluators, code-based or LLM-judge-based, fill any gap the built-in list misses. In the Contoso Outdoors agent playground eval (06:27), task adherence failed on several queries because responses summarized product features without pulling specific review data from searches.
Automated Red Teaming as a Pre-Ship Gate
The AI Red Teaming Agent attacks your endpoint without requiring a test dataset. Supply a callback function and a risk category list, and it generates adversarial inputs, fires them at the agent, and reports pass-through rates by attack type. Governance partners Credo AI and Saidot embed the same tooling to recommend which risk categories to test based on the registered use case.
“That’s the dimension that will make you break trust” — Mehrnoosh Sameki
In the AI Red Teaming Agent results walkthrough (17:22), 12% of attacks got through overall, including 21% of hate-and-violence attempts. A flip jailbreak, which reverses words in a harmful query to confuse the model, produced hateful output on the Contoso agent.
CI/CD Integration for Continuous Agent Evaluation
A GitHub Action and Azure DevOps extension, both available through their respective marketplaces, run batch evaluation against a test dataset on every commit. Compare mode adds a statistical analysis to confirm whether two agent versions differ significantly. In the GitHub Action compare-mode demo (24:53), adding Bing Search as a second knowledge source returned an inconclusive result, clearing the new version for deployment without manual review. Test datasets can be generated by a second agent in the Foundry playground to avoid hand-writing hundreds of queries.
Catching Regressions in Production: The Veeam Case
Veeam Intelligence is a support agent for Veeam Data Cloud for Entra ID, where a wrong answer during an identity outage can cascade into production failures. The team ran the Foundry GitHub Action as a regression gate: they disabled the knowledge-base tool call, pushed the commit, and watched the evaluation spin up. In the Veeam hallucination regression demo (35:07), similarity scores collapsed across most queries. Relevance, which needs no ground truth, failed on four questions — the signal that would have fired an alert in production.
Production Monitoring, Tracing, and Alerting
The Foundry monitoring dashboard tracks token usage, inference latency, and continuous evaluation metrics (relevance, task adherence, fluency, hate/unfairness) sampled at a configurable rate, such as 10 per hour. Each sampled run links to a full trace, so a low relevance score on a query surfaces the exact input, output, and LLM-judge reasoning in one click. Azure Monitor extends that data with Application Map dependency views, KQL access to raw evaluation scores, and alert rules that fire within five minutes when a score drops below a set threshold.
“I also live in KQL.” — Sebastian Kohlmeier
Notable Quotes
your AI Red Teaming Agent starts attacking your AI agent. Mehrnoosh Sameki · ▶ 16:52
That’s the dimension that will make you break trust Mehrnoosh Sameki · ▶ 12:26
I also live in KQL. Sebastian Kohlmeier · ▶ 47:43
Key Takeaways
- The Azure AI Evaluation SDK adds agentic metrics — tool call accuracy, intent resolution, task adherence — on top of standard safety and quality scores.
- The AI Red Teaming Agent requires no test data: point it at an endpoint and it generates adversarial inputs, reporting attack pass-through rates by category.
- The Foundry production dashboard samples evaluations continuously and fires Azure Monitor alerts within five minutes when scores drop below a configured threshold.