Hanchi Wang is a Principal Software Engineer Manager at Microsoft working on Microsoft Foundry, where he develops tools for observing, evaluating, and securing AI agents. His focus is the gap between an application that produces convincing answers and one whose behavior, accuracy, costs, and failures can actually be understood in production.
In 2024, Wang was an engineering lead in Microsoft’s Azure AI organization, working with Azure AI Studio and Prompt flow. A sales-analysis copilot he developed with colleagues translated business questions into SQL and generated visualizations; when it produced an incorrect chart, the failure demonstrated why successful tool calls do not guarantee correct results.
- Make agent execution observable. Wang uses OpenTelemetry-based instrumentation to capture model interactions, generated SQL, tool calls, and nested function execution, connecting local debugging with shared production telemetry.
- Evaluate the actual application. His application-specific evaluation compares generated SQL against ground-truth queries while measuring execution time, errors, and content safety. Testing multiple models against the same workload exposes concrete tradeoffs between answer quality and latency.
- Trace agents across clouds. For Microsoft Build 2026, Wang and Nagkumar Arkalgud co-owned a cross-cloud multi-agent observability project linking specialist agents on Microsoft Foundry, AWS Lambda, and Google Cloud Run. OpenTelemetry GenAI semantic conventions and W3C trace context combine their activity into one distributed trace, supporting routing diagnostics, latency analysis, and trace-based agent evaluation across different frameworks.
- Test safety before deployment. Wang advocates proactive AI red teaming to identify harmful behavior and security weaknesses before customers encounter them, treating safety as a production requirement alongside answer quality.
