The New Copilot Studio: What Changed, and Why Your Agents Suddenly Got a Lot Smarter

TL;DR: Copilot Studio has a brand-new build experience running on a new agent harness. One simplified screen, better models, reusable skills, memory, and an orchestrator with real tenacity. It is powerful enough that a working inventory management agent can be built with one sentence of instructions and one MCP tool. Here is what is different, what to watch out for on billing, and how to get results fast.

If you have built agents in Copilot Studio before, you know the drill. Lots of tabs. Lots of screens. Topics, nodes, triggers, and a whole lot of clicking to get from idea to working agent. And then, after all that work, you ask your agent to do something slightly ambitious and it shrugs at you and says it could not do it.

That is over. Microsoft has shipped a completely new way to build in Copilot Studio, and it is not a coat of paint. It is a different engine under the hood. In this post you will learn what the new experience actually changes, how skills and tools and memory fit together, what the billing change means for your budget, and then you will see a real example where one sentence of instructions produced a working inventory agent.

If you would rather watch the full walkthrough, check out the video here: The New Copilot Studio walkthrough

Or, if you want to see a fully functional agent built in 60 seconds check out this video: Build a Working Inventory Agent in 60 Seconds with Dataverse MCP

What is the new Copilot Studio experience?

It is a rebuild of the agent building experience at copilotstudio.microsoft.com. When you land on the familiar home page, you get a Try now option. Click it, the page reloads, and everything looks different. That is the new experience, and it runs on what Microsoft calls a new harness, which is their fancy way of saying this thing is fundamentally different under the hood, not just visually.

The good news for the nervous among you: it is a toggle, not a one way door. There is a new experience badge in the corner, and if you need to go back to work on one of your older agents, you click it, tell it why you are leaving, hit submit, and you are back in the classic builder. For now, you make a conscious choice about which side you are working on.

Why does the new build screen look so empty?

Because it is one screen. That is the whole build studio. In the old experience you had screens and tabs and places to jump around. Now you name the agent, set the icon (it is hiding behind the image itself, so click it, you are welcome), and everything else lives in one panel: model, skills, tools, knowledge, connected agents, memory, and instructions.

Here is the important part. Simpler does not mean it does less. It actually does more. The complexity moved from you to the orchestrator.

How is the new harness different from the old one?

Model choice is now yours

The first thing you will notice on the right is a model picker, and the list is different from the old agent tooling. You will see frontier class models in there. Opus 5 is the default at the moment. Your list may differ depending on the date and what your company has enabled, which is completely normal.

Choose deliberately, because the bigger the model, the more credits it burns. The most capable reasoning models are also the most expensive ones to run.

The orchestrator finally has tenacity

This is the real story. If you have used Copilot and then used Cowork, you have felt the difference. You ask Copilot to do something hard it tries hard and if it fails it says sorry, I could not. You ask Cowork and it just goes like a dog with a bone. That tenacity is what Microsoft brought to the new Copilot Studio harness.

Practically, that means your agents keep working the problem instead of giving up and asking you to try a different question. The old builder had a habit of saying I tried, it did not work, ask me something else. These new agents are far more robust than that.

Can you reuse your Cowork skills inside Copilot Studio?

Yes, and this might be the biggest quality of life win in the whole release. Skills are now first class citizens in Copilot Studio.

If skills are new to you, think of a skill as a recipe. It is a set of steps you want to be repeatable. Say every time you build a PowerPoint deck you want your brand colors, your fonts, your logos, and your standard intro slide applied. Instead of typing that into a prompt every single time, you package it once as a skill and call it forever. Skills can get far more complicated than that, but that is the right mental model to start with.

The reason this matters so much: skills use the skill.md format, which has become a universal standard. The same skill works in Cowork, in Copilot, and in other AI tools. Build it once, use it everywhere.

Sidenote: Skills are worth investing in. Lots of AI stuff comes and go but they seem to be core to a lot of platforms, learning them better is a good use of your time.

The packaging gotcha nobody warns you about

In Cowork you can find your skills under Customize. Open one and you will see the instructions plus any supporting files, like logo images. If you use the ellipses and click download, you only get the skill.md recipe file. Your logos and supporting files do not come along.

So package it yourself. Go to your OneDrive, into Documents, then Cowork, then Skills, and open the skill folder. Select everything in there, including skill.md and every supporting file, right click, and compress to a zip. Name it something sensible like brand-guide.zip. Then in Copilot Studio, go to skills, click upload (the upload button, not the tempting area next to it), pick your zip, and a few seconds later the skill is live in your agent.

No skill yet? Hit plus and create from blank. You give it a name, a description of when to use it, and the instructions. Pro tip: do not hand write the instructions. A real skill can run hundreds of lines. Ask an AI tool to build it for you. Tell it what you are trying to accomplish, let it interview you about your colors, fonts, and standards, and let it write the file. You can absolutely write it yourself, but why would you.

Do tools work the same way as before?

Mostly yes, with a slightly different interface. Hit plus on tools and you will find the same connectors you have been using in Power Apps and Power Automate for years, plus MCP servers like the SharePoint MCP server. Add a tool, click into it, and configure it.

The part people get wrong: name and description are not decoration, they are how the orchestrator decides when to call the tool. Do not leave a tool named Get items. Rename it to something obvious like Get inventory data. When your agent is misbehaving and not calling the tool you expected, nine times out of ten your name and description are the problem.

The rest is familiar. You choose whether the tool authenticates as the person using the agent or as the maker who built it. Inputs can be hardcoded values or left to AI, though the UI now makes you click into the field and choose New before the dropdown appears. If you leave an input set to AI, write a solid description of what you want generated, otherwise the agent will just ask the user for it at runtime. Outputs default to returning everything, and different tools expose different options, so poke around.

One more thing worth knowing: for something like sending email you often have a choice between a prescriptive connector action and an MCP server that just knows how to send email. Both work. Which one you want depends on how much control you need.

What changed with knowledge?

By default your new agent can search the whole web through Bing. A lot of the time that is not what you want, so turn it off. Then add the knowledge you actually trust: specific public websites, SharePoint, OneDrive for Business, or a file uploaded directly to the agent.

Here is the security detail to burn into your brain. If you upload a file directly to the agent, security is not honored. Anyone with access to the agent has access to that document. If you point at SharePoint or OneDrive instead, the agent uses the credentials of the person asking, so users only get results from data they already have rights to see. That difference will save you an awkward conversation someday.

What are connected agents and agent memory?

Connected agents let you bring other agents in as helpers. Your main agent can hand work to a child agent when it needs that specialty. This is how you build bigger solutions without cramming everything into one giant agent.

Memory is genuinely new here, and it is still in preview. The idea is your agent develops a sense of what you have done with it before, and that memory is tied to that specific agent. Imagine a reporting agent that asks what region you are in the first time and never has to ask again. If you have used memory in Copilot or ChatGPT, the concept is familiar. Early testing works well, but it is preview, so do not bet the business on it just yet.

What goes in the instructions?

Instructions are where you tell the agent its purpose in life. Your job is building PowerPoint decks or helping to manage inventory. When you get asked for a deck, call the brand guide skill. That kind of thing. It is also where personality lives, so if you want every response to include a dog joke, this is your moment. (I love a good dog joke or pun)

Aim for clear guidance rather than hyper specific micromanagement of every possible path. Tell it what it does, which tools and skills to use for what, and how to behave.

How do you test, evaluate, and ship it?

Preview lets you use the agent while you build. There is a great toggle up top that flips between the maker view, full of nerdy detail about which tool got called and what it was thinking, and end user preview, which shows what your users will actually see without the technical noise. Use the first one to debug and the second one to sanity check the experience.

Evaluate lets you set up test sets and run standardized tests against your agent. It is in preview, but it is there today. Monitor shows sessions, billing, and what is actually happening in production.

And do not forget to hit save. There is no autosave. Then use the publish dropdown to push to a demo website, embed in a web app, or (most likely) publish to Teams and Microsoft 365. Publishing takes a minute or two, then your channels light up and your agent shows up right inside M365 Copilot where your users already live.

What does this change about billing?

This one deserves your full attention. The new harness runs on Copilot credits, all the time.

• Building and previewing now costs credits. The old builder did not bill you while you were testing during the build phase. This one does. Every time the agent runs and thinks, that is credits.

• Your M365 Copilot license does not cover it. In the old model, depending on your licensing and scenario, some usage could be covered. On the new harness it is credits all the time.

• Model choice is a budget decision. The most capable reasoning models cost the most to run. Match the model to the job.

None of that is a reason to avoid the new experience. It is a reason to plan for it and to be intentional about which agents get the expensive model.

How much can it really do with so little? A one minute inventory agent

Here is the demo that makes the point better than any feature list. The goal was an agent to manage order items and vendors sitting in Dataverse. The build took about a minute, start to finish.

The entire build was two steps:

• One sentence of instructions: "You are an agent for managing order items and vendors for the Aug301 project using the Dataverse MCP tool."

• One tool: the Microsoft Dataverse MCP server. Connect, click add, done.

That is it. Ask it for the list of vendors and it comes back with the right four vendors: Amazon, Costco, Dell, and Target. Ask it to update Amazon to be valid through the end of 2027 and it writes the change back, because the Dataverse MCP reads and writes. Ask it to add a new webcam to the items table and it goes and finds the right table, pulls the schema, asks you for the cost and the vendor, and creates the record.

The kicker: the vendor column is a lookup. Nobody told the agent that it was a relational lookup or how to connect it. It figured that out on its own.

Should you ship a one sentence agent to production?

Please do not. It makes a spectacular demo, and it genuinely works, but one sentence leaves a lot to chance. Some things it will figure out. Some things it will not, and you will not know which until a user finds the weird path.

The production version of that same inventory agent took about five minutes. The difference was the instructions. Instead of one sentence, it listed the five specific tables the agent is allowed to touch (with the word only, so it never wanders into one of those other vendor tables), what to consider when creating orders, what to do when deleting orders, how to handle the status choice column, and general behavior rules.

Golden rule of agent building: you want the agent to think as little as possible. The more prescriptive you are, the less it improvises, and the less you have to worry.

Frequently Asked Questions

What is the new Copilot Studio harness?

The harness is the new engine underneath Copilot Studio agents. It replaces the old orchestration model with one that picks from newer frontier models, keeps working a problem instead of giving up, and supports skills, connected agents, and memory. You opt into it with the Try now button at copilotstudio.microsoft.com, and you can switch back to the classic builder at any time.

Does the new Copilot Studio cost more?

It changes how you are billed. Everything on the new harness runs on Copilot credits, including the time you spend building and previewing your agent, which the old builder did not charge for. M365 Copilot license coverage does not apply the way it sometimes did before. Bigger models consume more credits, so model choice is a budget decision.

Can I use my Cowork skills in Copilot Studio?

Yes. Skills use the universal skill.md format, so the same skill works across Cowork, Copilot, and Copilot Studio. To move one over, open the skill folder in your OneDrive under Documents, Cowork, Skills, zip the entire folder contents, and upload the zip in Copilot Studio. Downloading a skill from Cowork only gives you the skill.md file, so supporting files like logos have to be packaged manually.

Why does my agent ignore a tool I added?

Almost always because the tool name and description are too generic. The orchestrator reads the name and description to decide when a tool is relevant, so a tool called Get items tells it nothing. Rename it to something specific like Get inventory data and write a description that states exactly when to use it.

Is it safe to upload files directly as agent knowledge?

Only if everyone with access to the agent should also have access to that file, because directly uploaded files do not honor per-user security. If you point the agent at SharePoint or OneDrive for Business instead, it uses the credentials of the person asking, so each user only sees results from data they already have permission to view.

Key Takeaways

• The new Copilot Studio experience is opt-in through the Try now button, and you can switch back to the classic builder whenever you need to.

• The entire agent builder is now one screen, but it does more than the old multi-tab experience, not less.

• The new harness brings Cowork style tenacity, so agents keep working the problem instead of giving up and asking you to rephrase.

• Everything on the new harness is billed in Copilot credits, including build and preview time, and bigger models cost more.

• Skills built for Cowork work in Copilot Studio, but you have to zip the full skill folder yourself because downloading only gets you the skill.md file.

• Tool names and descriptions drive orchestration, so a vague tool name is the most common reason an agent ignores a tool.

• A working Dataverse agent can be built in about one minute with one sentence of instructions and one MCP tool, but a production agent needs prescriptive instructions and a human review.

• Filling in Dataverse and SharePoint description fields is the single highest value prep work you can do today for the agentic era.

Need a Hand?

Everything above comes out of real client work, whether that is training teams to build this stuff, building it for them, or co-building it alongside them. If you want help building agents like these, sorting out the difference between Copilot, Cowork, the old Copilot Studio, the new Copilot Studio, and Foundry, or just getting your team up to speed, the team at PowerApps911 can help.

Reach out by clicking the Contact button or check out the upcoming live Copilot Studio class over at training.powerapps911.com. Let's build something awesome together.