Join me exploring Microsoft Copilot through each letter of the alphabet. This week, S is for Studio.

I’ve been scared of Copilot Studio for ages. It sounded like “code” and beyond me. But S had to be Studio, so I rolled up my sleeves appropriately. 

Caveat before we start: I’m writing this in the middle of a UK heatwave, it is roughly five hundred degrees in my house, and if this post is slightly unhinged in places, that’s why.

Ignoring my own advice

Having worked with Copilot for a while now, I thought it would be able to spot patterns in my work and come up with some good ideas for an agent I should build. It did, but I ignored all of them and pursued my own idea instead. I wanted an agent to help me with my weekly work review.

I do a version of this weekly review ritual at home too, which I’ve built a Gemini gem for (another blog, another day) but at work it’s been a manual shuffle through checklists and documents. Time an agent took that off my hands.

I wanted the agent to help me: 

• Celebrate wins
• Close loose threads
• Make sure i’ve got time set aside next week to achieve my goals 
• Check for calendar clash madness to sort out next week

Quick side note: you can build a simple agent straight from “New agent” in the Copilot nav, without going anywhere near Copilot Studio. But that’s not technically Studio, and this is an S post, so I went and found the real thing.

Finding Copilot Studio (embarrassing bit)

I had to ask Copilot how to find Copilot Studio. Facepalm. In my defence, it’s pretty hard to find if you don’t know where you’re looking. It’s at copilotstudio.microsoft.com, and you need a licence for it. 

A pop-up asked if I wanted the new experience. I said yes, mostly because I’d never used the old one to compare. It looks nice, nothing scary so far. There are two options sitting side by side: agents and workflows. (More on Workflows in a future post.) 

Building the thing – the scary part (or not?)

When you start building, the first job is giving your agent a name and some instructions. Handily, I’d already had a chat with Copilot to help me compile the instructions, so I had some pre-written and ready to paste in.

Then there’s a whole side panel of options. Mostly self-explanatory, but here’s what each one actually does, in the order you’ll hit them:

• Model. Which model runs your agent (more on that in L is for LLM)
• Microsoft IQ. Needs to be on if you want the agent looking at all the “stuff” you’ve been working with. 
• Skills. This could be a whole other blog post – maybe my next one. It’s how you get the agent to do a task in a specific way repeatedly. I’ve skipped it this time.  
• Tools. This is where the properly fancy, slightly scary stuff lives, like connecting your agent to other things in the big wide out there. 
• Knowledge. Where you tell your agent what to look at, which doesn’t have to just be internal sources but can also be public websites as well.
• Connected agents. You can connect agents to other agents! Mind blowing. Filing that away for a future post, not touching it today.
• Memory. A preview feature. I want this agent to get to know me and spot patterns over time (see P is for Personalisation), so I turned it on.

I was expecting to do a lot more heavy lifting to build an agent. But actually, that was it! Once I’d written my instructions and chosen my toggles, I could then save it, preview and publish it. 

Preview was useful, it let me see how the agent would behave in real life when prompted. 

You can also choose whether to publish just for yourself or make it available to others in your org (not the whole world, thankfully).

The mindset shift, not the tool, is the hard part

I kicked off my new agent with a suggested prompt and watched it ask for some permissions and have a think.

While I waited, I had an annoying thought. I’d just built an agent to do something I could already do by opening Copilot chat once a week. Which seemed a bit like a waste of an agent. 

An agent should actually do things for me. I’d basically ignored all my learnings from writing the rest of this series and not done the proper thinking about what outcome I wanted. I’d done the classic jump straight in and see what I can build. 

That’s the real mental gear change with all of this: not “how do I use this tool” but “would an agent, doing this on its own, actually save me any time / brain space.” That’s the point that a prompt stops being a prompt and becomes an agent.

So the scary part isn’t Copilot Studio itself at all. The scary part is curating a totally new mindset shift towards how work gets done. I’m still working on that.  

Take two: the spot-what’s-dropping agent

Not easily beaten, I tried again, this time with one of Copilot’s ideas; a “spot what’s dropping” agent. Just as quick to build, save, and preview, and Copilot wrote the instructions for me again based on what it had already noticed about how I work.

This one seemed more in the ‘agent’ direction. In the preview it came back with a formatted table of things that seem like they are slipping, suggested owners, and urgency – including one thing that had been bouncing round the back of my brain as an unresolved open loop for days.

This still felt a little bit like ‘useful prompt’ territory though. It would be much more useful if the agent came to me with this information on a schedule, without me lifting a finger. Here’s the but – running the agent automatically is a Workflow feature, not an Agent feature, in Copilot Studio. Two different things living in the same tool. (That’s exactly what my W post is for.)

Publishing, or: where did it go?

Publishing takes a while. When it finished, I couldn’t find my own agent anywhere.

Gotcha #1: once you’ve published, the share link doesn’t take you to Copilot. It takes you to Teams, and turns your agent into a Teams bot whether you asked for that or not. Fine if that’s useful to you, but it caught me off guard and wasn’t what I was expecting. (Even though it basically told me that’s where it would go…) 

Gotcha #2: To see your agent inside Copilot itself, you have to publish to M365 Copilot specifically. You need to tick a separate box for that.

As a first attempt, I’m pretty happy with this agent. Even if it doesn’t run on a schedule yet, I can still use it to help me spot and close open loops. It feels like at least the start of something I can iterate on over time, and it’s freeing up some mental capacity, so I’ll take that as a win. 

TLDR

Copilot Studio itself isn’t scary at all! The interface is user-friendly and straightforward, and Copilot will happily write your instructions for you if you ask nicely first in chat. The actual hard part is the mindset shift: learning to spot the moments where you need something running for you in the background, rather than a prompt you remember to ask each week. 

I’ve built an agent I’m happy with. I still haven’t fully cracked agentic thinking. Getting it to run on autopilot is a Workflow problem, which is next.

Have you built anything in Copilot Studio yet — and did you get further than a glorified weekly prompt? Tell me in the comments, ideally from somewhere cooler than my house right now.

Useful links:

Copilot Studio: https://copilotstudio.microsoft.com/ Get started with Copilot Studio: https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-get-started

And if you’re ever in London and fancy hanging out with other Copilot-curious folk: https://mccug.co.uk/