When people talk about AI agents, everyone argues about the same thing: which model is smarter. GPT, Claude, whatever came out last week. It is an entertaining conversation and, for most companies, fairly irrelevant.
Because in the meantime Microsoft changed something else in Copilot Studio, its tool for building agents. Something with an unfriendly name, harness, that decides how an agent works, what it can do and how much it costs.
And no, it is not a detail for the IT department.
The boss nobody sees on the org chart
Think of an agent as a very smart employee. The AI model is its brain. But a brilliant brain is not much use if nobody tells it how to organize itself, which tools are at hand and what to do when something goes wrong.
That is the harness: the agent's way of working. Microsoft describes it as the layer that sits between what you configure and the model, deciding when to think, with what information and which tool to use next.
Two agents with the same brain and a different harness behave like two people with the same university degree: one follows the manual to the letter, the other gets a goal and figures it out alone.
The model brings the intelligence. The harness decides what to do with it.
Three ways of working, three personalities
Today Copilot Studio offers three:
| How it works | What it is for | |
|---|---|---|
| Standard | Follows the script you write | Predictable support: help desk, FAQs |
| GitHub Copilot | Gets a goal and builds the plan itself | Long processes with several steps and documents |
| Copilot Chat | Answers with your company's content | Helping people find information without leaving Copilot |
Standard is what, until June, we simply knew as Copilot Studio. It does exactly what you asked, no more, no less. Reliable and somewhat rigid.
GitHub Copilot is the big news. Microsoft's example is an accounts payable process: the agent reads invoices, matches them against purchase orders and sends exceptions for approval. If something fails along the way, it looks for another way to solve it. It can also create and edit Word, Excel, PowerPoint and PDF documents, and remember things from one task to the next.
Copilot Chat is the simplest: the classic welcome agent that answers a new employee's questions using what is published in SharePoint.
The name that will scare someone at your company
Yes, one of them is called GitHub Copilot. No, your data does not go to GitHub.
Microsoft says so explicitly: it shares technology with GitHub Copilot, but company information does not go through that service, and the same privacy and compliance rules as always still apply.
Have that answer ready. In some committee, someone will read "GitHub" and raise a hand.
The meter starts running before you get in
This is what will weigh most on budgets, so let's get to the numbers.
Copilot Studio does not charge per user or per agent. It charges in Copilot Credits, a kind of token that gets spent every time the agent does something. There are two ways to buy them:
| Option | Price | Detail |
|---|---|---|
| Monthly pack | USD 200 for 25,000 credits | Whatever you do not use that month is lost |
| Pay as you go | USD 0.01 per credit | Billed at month end, through Azure |
Companies that prepay a full year can get up to 20% off.
In the traditional model, spending is fairly easy to calculate, because each action has a fixed price:
- A scripted answer (one you wrote): 1 credit
- An AI-generated answer: 2 credits
- An action, like creating a record or sending an email: 5 credits
- An answer that searches the company's internal data: 10 credits
A simple example: with a USD 200 pack, an agent that only gives AI-generated answers covers about 12,500 answers a month. If each answer also searches company documents, the same pack goes much less far.
There is good news hidden in there: if employees have a Microsoft 365 Copilot license and use the agent inside Teams, SharePoint or Copilot Chat, much of that consumption is not charged.
In the new model, the math is different. Billing starts as soon as you begin building the agent: testing, tweaking and testing again also spend credits. And there is no fixed rate per answer, because consumption depends on how much the agent "thinks", how many tools it uses and how long it takes to solve the task. Microsoft does not publish a price per task: it offers a calculator to estimate it.
The "let's play around and see what happens" phase is no longer free.
And Microsoft itself admits something you rarely read in a vendor document: a more capable agent can overcomplicate a simple problem, taking longer and costing more without delivering a better result. Using the most powerful harness to answer how many vacation days you have left is hiring a senior consultant to answer the phone.
In the traditional model you pay per answer. In the new one, you pay per effort.
There is no upgrade button
If your company already has agents running, the question is obvious: how do I move them to the new model?
You don't move them. You rebuild them. There is no automatic conversion from one way of working to the other.
The good news is that Microsoft is not pushing either: its own guidance says that an agent that works well does not need to move just because something new showed up. Switching makes sense when there is a concrete benefit, not because it is trendy.
The point
Choosing how your agent works is no longer a technical decision. It is a business decision: it defines how predictable it is, how far it can go and when it starts costing money.
In the next article we will compare the classic model with the new one, with simple criteria to know which fits each case.
If you don't want to miss it, subscribe to the macareno.net newsletter.
Sources
- Harnesses in Copilot Studio, Microsoft Learn
- Choose a harness for your agent, Microsoft Learn
- Billing for agents powered by the GitHub Copilot harness, Microsoft Learn
- Billing rates and management, Microsoft Learn
- Microsoft Copilot Studio licensing guidance, Microsoft
