Use evals before changing prompts
Treat prompt changes like code changes: measure the behavior before deciding whether the edit helped.
Ai-Engineering
Keep RAG citations tied to retrieved chunks
Citations should point back to the retrieved evidence, not to sources the model guessed after answering.
Observability for Agents with OpenTelemetry, Aspire, and Application Insights
How to make Microsoft Agent Framework behavior visible in production with OpenTelemetry traces, token usage, model calls, tool calls, Aspire local debugging, and Application Insights monitoring.
Use CSnakes when Python is a function call, not a service
Embed a small typed Python capability in .NET when an HTTP service would be more boundary than value.
Expose a Python FastAPI AI tool as an Aspire service
Keep Python AI code behind an HTTP boundary and let Aspire wire it into the local system.
Keep model name and deployment name in config
Do not hard-code model routing decisions into application logic.
Keep vector search filters separate from semantic ranking
Use filters for eligibility and ranking for relevance instead of blending both concerns.
Run Python in-process from .NET with CSnakes
A quick tip for embedding Python directly into a .NET app with CSnakes when you want a small slice of the Python ecosystem without a separate API service.
Separate prompts from authorization
Do not let prompt instructions decide what data or actions a user is allowed to access.
Use approval for side effects, not for every tool call
Reserve human approval for actions where a wrong call would actually matter.