Copilot Studio has been around for a while now, originally Power Virtual Agent before the Copilot Revolution, and as chatbot builders went, it was ok.
But it never really felt right for me, it topics felt cumbersome (and this is from a Power Automate guy), and I never got the results I expected, and some how it felt super complicated. The simple truth was it was an old fashioned chatbot builder with generative AI bolted on, and it could never deliver compared to GitHub Copilot, ChatGPT, Claude, or Gemini.
The thing is it wasn’t just the UI or design, it was built on a different engine (check out this great article explaining more perspectives.plus), and Microsoft knew it was on the path, so migrated the backend and created a new front end, and that’s what I want to talk about today.
Hopefully you have seen the new Copilot Studio, if you haven’t go to your preview environment, or slapped .preview in your Copilot Studio url (copilotstudio.preview.microsoft.com)you would have seen the new Copilot experience.
- How its different
- What’s New
- What’s missing
- Is it any good
1. How its different
Apart from the engine, the main difference is the removal of topics. Topics were the old way of building out predefined responses from your agent. It included question/answers and actions like API calls. Topics have been around since Power Virtual Agents, before the LLM era. Copilot Studio bolted on some AI to help channel requests to topics, but they were more like Power Automate then GitHub Copilot. They have been replaced with Skills, which I will go into more detail about later.
There is also some new features but this is main different, as it represents a fundamental change, from:
- The maker trying to decide on actions and AI deciphering the user to find right topic
to:
- The agent decide on the right actions to take
There has also been a change in the deterministic side, with Agent Flows been replaced with Workflows (a new UI with some new AI features), but I will go into that in another blog. Additionally both the UI and UX are much improved, options are easier to find and the preview/test is much better, being with not only full screen but showing the steps/reasoning the agent is making.
2. What’s New
The new builder focuses around 6 main areas.
The models are the same, the instruction the same, but the UI is now simpler and feels more like a modern LLM.
Microsoft IQ
If you see IQ, it means Microsoft’s way of grounding data for your agent. And by that I mean the agent gets access to your organisations data, currently that covers.
Work IQ – All of your data across Office etc, your emails, chats, OneDrive files, the agent will have access to it all to add to its context when you ask it anything
Fabric IQ – All of your business key data.
Foundry IQ – Knowledge sources created in Foundry.
Work IQ in particular is very cool, you don’t need to set anything up and your agent instantly has on-behalf access to the current users data.
https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/use-microsoft-iq
Skills
If you don’t know about skills you really need to, they are the best thing Anthropic have created (nope not MCP, I’m not a big fan, but that’s a conversation for another day).
You can think of Skills as optional prompts and context that the LLM can decide to use. They are markdown files and support files (like scripts, templates, or knowledge docs), and are optional, so only used when the LLM thinks its needed (you can see how they replace topics). So in a nutshell we get:
User request & instruction – LLM asks for skill – request & instruction & skill
Skills are super cool, as a basic one is just a simple instruction file that is used for specific use case, just text and you get the benefits of lots of complex workflow logic.
They deserve a blog all for themselves (watch this space), but needles to say they are one of the best tools Claude Code, Codex, and GitHub Copilot uses, so its great to see them in Copilot Studio (there are also thousands already available for you to use/get inspired by, Anthropics GitHub has some good ones.
https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/skills-overview
Tools
Same as classic Copilot Studio, these are your connectors, MCP’s, and Workflows (old Agent Flows). Just like your Skills they are available for the LLM to decide when to use them.
https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/tools-overview
Knowledge
Again like tools the same as classic Copilot Studio. Here you can add knowledge for the LLM to use to answer questions or make decisions
Connected Agents
Deja-vu, again nothing new, here will can link child agents and other agents to handover tasks. This was critical for the classic Copilot Studio to get ring fenced context. But Skills can do this now, so I think it will become really useful when you have a large ecosystem of agents. That way they can seamlessly pass you to the right agent if you are off topic or just lost.
Memory
Memory is a simple toggle, and its kind of in the name, the agent can now remember previous sessions. Users can see all of their memories and manage them. I’m excited to see how this works, as context is a fine-line to walk, too much and it can have negative impacts on your requests. Though on the flip side it will be able to kind of learn, with it responding in the way you prefer and understand your requests better (come on that’s cool isn’t it 😎).
https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/memory-overview
3. What’s missing
The new Copilot Studio is in preview, so I expect most of these to be added over time, but as of now:
- Access to the solution, you cant see the solution area in the environment, so if you want to export you need to got to the classic or Power Automate/Power Apps
- All required components are not automatically added to the solution when you add the agent
- All of the Knowledge sources, you are currently limited to External sites, SharePoint, and OneDrive
- Autonomous Agents, though this is because the change in strategy, with the new Workflows taking on the trigger of agents. I think this is the right way to go, so I suspect Autonomous Agents will disappear for good
4. Is it any good
So is it any good, Microsoft doesn’t have a good record for UI updates (cough Power Automate V3 launch), but in this case, hell yes.
This is what Copilot Studio should have been from day one, it now feels and acts like other LLM agents. It no longer feels like the old legacy Power Virtual Agent, it is something new and modern.
The results are smarter/better, you can do more, and you can build a lot quicker. I personally will not be touching the classic UI again, there is no migration path (different engine), and the results are just so much better (the only blocker is the knowledge sources, but in that case I would rather just wait). And with the pace (and total focus Microsoft has on Copilot Studio) I excited to see new feature updates soon.
So if you haven’t used it, you need to try, trust me you wont look back.
답글 남기기