From the Microsoft Power Platform blogs: Agent evals; Canvas app development; Evaluation experiencess; Copilot Studio code interpreter

In this review of the Power Platform blogs:

• Who evaluates the evaluators? The data science behind agent evals
• AI-assisted canvas app development using external tools in Power Apps
• Copilot Studio evaluations
• Streamlining SharePoint file analysis with Microsoft Copilot Studio code interpreter

Who evaluates the evaluators? The data science behind agent evals

On the Microsoft Copilot Studio Blog, Dikla Dotan‑Cohen wrote that Copilot Studio offers many ways to improve AI agents, including adding knowledge sources, refining prompts and adjusting models. 

But how can you tell if those changes actually make the agent better? How do you spot problems before users do, and how much can you trust the results?

Dotan-Cohen stated that data scientists face the same challenge when developing AI models. Before releasing a model and after every significant change, they test performance, track metrics, compare versions, and look for regressions to answer one key question: Did this change improve the model?

She noted that when Microsoft built evaluation features into Copilot Studio, it took the same approach. The goal was to determine whether the evaluation process could reliably answer the question: Is it better? She stated that this is the quality question.

To answer that question, Dotan-Cohen explored three core areas of AI evaluation quality.

AI-assisted canvas app development using external tools in Power Apps

On the Inogic blog, Sam Kumar wrote that canvas apps have long been popular because they let users build applications quickly with a low-code, drag-and-drop approach. But Microsoft recently added support for external AI-powered development tools for canvas apps. This gives developers and business users the ability to use AI-assisted tools outside of the traditional Power Apps Studio environment.