Agentic Development Is Only as Good as Its Feedback Loops

Truly agentic development requires one thing: Feedback loops ➰.

Many of them.

Take refactoring a Power BI semantic model from scratch. The obvious feedback loop is: "Do measures still return the same results after refactoring?"

That only proves you didn't break the model. It does not prove you improved it.

So you add another loop: "Did the model structure improve?" Fewer many-to-many relationships. Cleaner dimensions. Better granularity. Proper date table or calendar. Clearer separation between facts, dimensions, and measures.

Then another: "Did the measures improve?" Less duplicated logic. Better naming. Cleaner DAX. More reusable base measures. Fewer calculated columns and tables pretending to be metrics.

Then another: "Did performance improve?" Faster queries. Lower memory usage. Cleaner storage engine / formula engine balance. Fewer unnecessary transformations at report time.

I believe that the future is in giving agents large and small feedback loops so they can improve safely.

Take SemanticOps, as an example, where I execute on this idea:

• Test runner - done
• Performance profiling - done
• Static analysis for DAX - coming soon
• Static analysis for Power Query - coming soon too

Same idea in PBIR CLI, developed by Kurt and myself. When AI creates or changes a visual, it gets immediate feedback: Too small? Too large? Overlapping?

Agentic development is only as good as the feedback loops you give it. Good feedback loops let you elevate even smaller, cheaper models.

That is why I believe the future is loops ➰.