Diving into the details
The solution from Part 1 of this blog post isn’t the only way to develop with agents. You can also work with the Claude Code extension, for example, as well as other solutions. I went with this approach because I’m personally at home in the Microsoft universe and wanted to use that stack.
In this blog I want to build on the experiences from the first one. The post is meant to showcase the features of the development environment. Just to be clear about the setup up front: I’m driving Claude’s agent through GitHub Copilot, so some of the features here come from Copilot itself (model choice, modes) while others come from the Claude side (skills, hooks, plugins). And again the disclaimer: you should never try out agents on business-critical data — only on non-critical demo data.
Let’s get started
A new agents window has been available for a short while now. You can open it in Visual Studio Code via this button. The look and feel is reminiscent of web-based LLMs.
That gives you the following window.
One thing worth highlighting: the new Agents window in Visual Studio Code doesn’t work directly on your active branch. Instead, it spins up a dedicated worktree inside your local Git repository — an isolated working copy where the agent makes its changes. Once you’re happy with the result, it’s on you to commit those changes back into your local main branch.
If that workflow feels unfamiliar, you can always fall back to the classic mode and keep working the way you’re used to. But the worktree approach has a real advantage: your current solution stays completely untouched while the agent experiments. Nothing gets overwritten, nothing breaks — you review the changes on your own terms and only merge them in when you’re confident they’re right. The screenshots cover the new Agents window, but are similar in the old experience.
Choice of Model
Within GitHub Copilot you can select different models. In this example, Claude Sonnet 4.6 is currently selected.
But you also have other models to choose from, like Claude Opus or GPT. Not every use case needs the most modern model, such as Opus or GPT-5.5. Don’t forget that, at the end of the day, the most modern models are also the ones you pay the most for.
You can also define the thinking effort. In this example it’s set to Medium. You can also turn reasoning off entirely.
Modes
Within the new agent window you currently only get Agent mode. In the classic chat view you can switch between Ask, Plan, and Agent.
- Ask only explains and answers like you know from the ChatGPT or other web interfaces.
- Plan explores your Power BI directory and produces a reviewable plan before any code changes start.
- Agent plans the development steps, edits across multiple files (TDML, JSON), runs commands (for example with the MCP Server), and iterates until the task is done.
Custom Agents
Important to know you can also see your own agents here. The image below shows custom agents you can download from the repos of Kurt Buhler and Rui Romano. Check each repo’s license before any commercial use, since the terms vary.
You can also build agents like these yourself — I’ve run various experiments together with GitHub Copilot. For example, the powerbi-planner is a Power BI planning agent.
Skills
Another feature is what are called skills. A skill gives an agent domain expertise, best practices, commands, and workflows. Here you can see various examples — from Rui Romano and Kurt Buhler, as well as ones I wrote myself with GitHub Copilot.
Instructions, Hooks and Plugins
Beyond that, there are Instructions, Hooks, and Plugins:
- Instructions are general guidelines and rules that directly influence how an agent carries out development tasks.
- Hooks are commands or scripts that run automatically at a specific point in the workflow.
- Plugins bundle related capabilities — skills, custom agents, commands, hooks, and MCP servers — into a single installable package. Kurt Buhler, for example, has written such plugins.
Links
If you want to dive right in, here are a few essential links for working with Power BI Agentic Development:
Rui Romano’s Power BI Agentic Plugins: An incredible, hands-on GitHub repository. It provides pre-built “plugins” (skills, agents, and connectors) designed to turn GitHub Copilot into a specialized developer for Power BI and Microsoft Fabric.
Data-Goblin’s Power BI Agentic Development Marketplace: A highly useful plugins by the one and only Data Goblin Kurt Buhler. It features a collection of plugins, skills, and hooks (like Best Practice Analyzer rules and DAX validation) specifically focused on improving how coding agents interact with Power BI.
Use the Agents window (Preview): tutorial how to use the new agents window.
Screenshorts: “Used with permission from Microsoft”