In June Microsoft introduced a rebuilt Copilot Studio: agents that receive a goal, build their own plan and work with documents without anyone drawing the path for them. The demo is flawless and the temptation is obvious: move everything to the new model.
Before doing that, it is worth reading what Microsoft says in its own guidance. The classic Copilot Studio is still supported, still gets new agents and, for certain kinds of work, Microsoft itself recommends it over the new one.
Two Copilot Studios living under the same roof
Today there are two ways to build agents in the same tool.
The classic one works from a script. You define the topics, the questions and the paths, and the agent follows them. If someone asks something that was not planned, the agent does not improvise.
The new one works from goals. You give it a task, for example reviewing invoices and matching them against purchase orders, and it decides the steps, uses the tools it needs and, if something fails, looks for another route.
The two coexist and neither converts into the other. A classic agent cannot be moved to the new model with a button: it has to be rebuilt.
| Classic | New | |
|---|---|---|
| How it works | Follows the script you wrote | Builds its own plan from a goal |
| How predictable it is | Very predictable | May take different paths each time |
| Documents | Processes files in small pieces | Creates and edits Word, Excel, PowerPoint and PDF |
| Memory | Does not remember between conversations | Remembers preferences and corrections |
| When billing starts | When you publish | From the moment you start building |
| How it bills | Fixed rate per action | Based on how much the agent works on each task |
| Maturity | Complete and proven | Still missing some classic features |
First advantage: you know exactly what it will answer
In a help desk, a vacation request form or customer service, the agent's creativity is a risk. Nobody wants the HR bot to interpret the leave policy with artistic freedom.
The classic one answers what you taught it. That is why Microsoft recommends it when the task is short, structured and needs predictable behavior. It also answers faster, because it does not stop to reason through every step.
Second advantage: you can do the math with a calculator
Copilot Studio bills in Copilot Credits. A monthly pack costs USD 200 for 25,000 credits, and with pay as you go each credit costs USD 0.01.
In the classic model each action has a fixed price:
- Scripted answer: 1 credit
- AI-generated answer: 2 credits
- Action, like creating a record or sending an email: 5 credits
- Answer that searches company data: 10 credits
With that, someone in finance can estimate the monthly spend before starting. And 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 there is no rate per answer. Consumption depends on how much the agent reasons, how many tools it uses and how long it takes, and the meter runs while you build and test it. Microsoft does not publish a price per task: it offers a calculator to estimate it.
In the classic model you know what an answer costs. In the new one, you find out at the end of the month.
Third advantage: it is already finished
The classic one has years of use, proven connectors and is available on all the usual channels. The new one arrived in June, and Microsoft still advises checking that the features and channels you need exist before choosing it.
For an agent that already serves hundreds of people a day, swapping something that works for something still being completed is a hard bet to justify to the committee.
Where the new one wins comfortably
The new model exists for a reason. When the task is long, involves several systems and requires decisions along the way, the classic one falls short: every unexpected case forces you to write more rules, and the script ends up looking like a maze.
The new one handles that kind of work end to end. It reads an entire spreadsheet, cross-checks data with another system, generates the report in Word and, if a step fails, tries another path. Microsoft's most advanced reasoning models are only available there.
Microsoft's examples are preparing customer meetings, compliance audits and accounts payable reconciliation. Work that today takes hours of someone's time.
The expensive mistake: using the new one for everything
Microsoft's own guidance warns that a more capable agent can overcomplicate a simple problem, take longer and cost more without delivering a better result.
Putting the new model to work answering the cafeteria's opening hours is paying for reasoning nobody asked for, with a bill that grows from the first day of testing.
The rule to avoid overpaying
Three questions settle most cases:
- Is the task short and repeated the same way every time? Classic.
- Is it long, does it use documents and require decisions along the way? New.
- Do you already have a classic agent that works well? Leave it where it is. Microsoft says explicitly that it does not need to move just because another option exists.
Microsoft also proposes a combination: a new agent working overnight to organize scattered spreadsheets and documents, and classic agents serving people during the day using that cleaned-up data. Each one does what it does best.
To get the next analysis, subscribe to the macareno.net newsletter.
Sources
- Choose a harness for your agent, Microsoft Learn
- Harnesses in Copilot Studio, Microsoft Learn
- Classic vs. new agent experience, Microsoft Learn
- Billing rates and management, Microsoft Learn
- Microsoft Copilot Studio licensing guidance, Microsoft
