Copilot Studio: agents with their own account and a memory

Every Tuesday, an agent prepares the week's budget review. It gathers the forecasts, compares them with the explanations received and chases the budget owners when a variance isn't explained. That's the example Microsoft came up with to present the new Copilot Studio agents, announced on 15 October 2026.

What Microsoft is announcing

Copilot Studio is Microsoft's tool for building AI agents: assistants that read, prepare and follow up, inside Microsoft 365. On 15 October 2026, Microsoft presented three new features for agents that take charge of a process shared by a team.

• Agent accounts. The agent works with its own credentials.
• Process memory. The agent keeps the context from one cycle to the next.
• Triggers. The agent starts on its own, on a signal or on a set schedule.

The underlying idea: agents that belong to the organization, not to the person who created them. They stay in place when that person changes jobs or leaves.

An agent with its own account

This is the most important new feature for executives. The agent works in Microsoft 365 with its own credentials, as a person would. Not with a user's credentials, nor with those of the person who built it.

In practice:

• you reach it in Teams or Outlook, to ask it a question, give it a task or pass it some information;
• you can mention it in a comment in Word, Excel or PowerPoint;
• it keeps working between two exchanges.

It acts with limited permissions, approved for it. And what it does is attributed to it. You can see which actions come from the agent, and which come from people. The process owner stays in control of its scope.

On the administration side, administrators can limit its access, add controls and manage its lifecycle from the Power Platform admin center. For the coming months, Microsoft announces further access rules, along with quality, cost and lifecycle controls specific to these agents.

A memory that serves the whole team

Second new feature: process memory. From one cycle to the next, the agent keeps the users' preferences and the context of the process.

The interesting point: what one person contributes becomes a resource for the whole team. You no longer have to explain everything again each time you give it a task.

Whoever builds the agent stays in control. They decide, through instructions, how this context is used. They choose the documents and rules that are authoritative. Stored memories can be viewed and deleted.

Agents that start on their own

Third new feature: triggers, for agents built on Copilot Studio's new agent engine. An outside signal, such as an incoming email, or a schedule set in advance starts a work cycle. Without anyone having to ask.

Microsoft also announces, with no date, agents that will have their own computer in the cloud, with a browser. And, in the coming months, recommendations drawn from the agent's activity. They are meant to help turn repeated steps into automated processing, spot ambiguous instructions and choose a model better suited to the task.

The finance department example

Let's go back to the example Microsoft came up with. Every Tuesday, the agent prepares the week's budget review. With its own credentials, it gathers the planning data. It compares the forecasts with the explanations received. It asks the budget owners to explain the variances.

The finance team adds context by mentioning it in Teams or replying to it in Outlook. The agent produces a summary of the figures, the explanations and the gaps. The decisions, though, go back to finance: revise the forecast, ask for more information or approve the plan.

In the next cycle, the agent starts again with the review requirements, the open questions and the relevant context. The team spends less time piecing together past discussions. Finance stays in control of allocations and spending.

What this means for you

For executives, the real news is neither the memory nor the triggers. It's traceability. An agent with its own account, its own permissions, and actions attributed to it, is an agent you can audit. You know what it did. You know what the people did.

It's the principle I applied in a whistleblowing channel: security lives in the data and in the permissions, not in the interface. I told that story in hiding a button does not protect the data. An agent with its own account is configured the same way: through its permissions.

Triggers address another problem I know well: a deadline watched only by humans ends up missed, by statistics. An agent that starts on its own every Tuesday doesn't depend on anyone's memory. That's the subject of the article on regulatory deadlines.

Three questions to ask yourself before handing a process to an agent:

• Who owns the process? The agent acts, but someone sets its scope and answers for the decisions.
• What permissions should it get? As few as possible. An agent with its own account is set up like a new hire: it only gets access to what it needs.
• What should it remember? A memory shared by the whole team is useful. It also needs a clear rule: what is authoritative, and what must be deleted.

For now, these features come through Frontier, the Microsoft 365 early access program. They're recent features that may still change, available depending on your license and your administrator's settings. Microsoft announces neither a general availability date nor a price. The right time to prepare is before general availability: choose the process, its owner and its rules. The AI agents I build are described on the page A Copilot AI agent that works for your teams. A real example: the case study An AI agent that answers questions on the workshop's orders.

Based on Microsoft

Based on: Delegate shared business processes to agents that are more persistent and proactive with Microsoft Copilot Studio, Microsoft, 15 October 2026, https://www.microsoft.com/en-us/copilot/blog/copilot-studio/delegate-shared-business-processes-to-agents-that-are-more-persistent-and-proactive-with-microsoft-copilot-studio/.

Article published on the Copilot Studio blog by its product vice president. The sentence that sums up what's at stake for executives: "work can be attributed to the agents, while process owners retain control over its scope".