Copilot now lets people build their own agents: an assistant that answers from a chosen set of documents, follows set instructions and, in Copilot Studio, can take actions in other systems. It takes minutes and needs no developer, which makes it one of the more useful things Copilot does.
Because it is that easy, a business can go from no agents to dozens in a quarter, each built by someone different, connected to something different and shared with people the builder chose.
We are a Microsoft Solutions Partner for Microsoft 365 Copilot, and Cardonet is Certified in the Claude Partner Network. What changes when a tool can act sets out the general case for agents; this is the Microsoft version.
Two kinds of agent
Agent Builder, inside Copilot, makes simple agents, while Copilot Studio makes fuller ones.
| Agent Builder | Copilot Studio | |
|---|---|---|
| Built from | Instructions plus a few knowledge sources, such as a SharePoint site or a set of files | Connections to other systems |
| What it does | Answers questions and drafts | Carries out actions, runs flows and can start work on its own when something happens |
Although the decisions below apply to both, they weigh more heavily on agents built in Copilot Studio.
Six decisions before people build
1. Who may build and who may share
Letting people build agents for their own use is low risk and worth allowing. Sharing one with a team or the whole business is different, because everyone it is shared with now relies on it. Before people start, decide that anything shared beyond its builder is reviewed first, and that Copilot Studio, with its actions and connectors, is limited to people who have been shown how to use it safely.
2. What each agent may know
An agent built on a SharePoint site or a folder answers from that content. Check that the content is what the agent's audience should see, and that it is current, because an agent answers from a superseded policy as confidently as from a current one.
For every knowledge source an agent is given, the checks in What Copilot can already see apply, and where a source holds sensitive files, sensitivity labels before Copilot covers what to label first.
3. Whose access it acts with
Some agent connections can be set up to run with the builder's own account rather than the account of whoever is using the agent, in which case anyone using it acts with the builder's access.
4. Which connectors and actions are allowed
Although Copilot Studio agents can connect to hundreds of services, Microsoft's data policies let an administrator decide which connectors may be used together and which may not be used at all. Start narrow: internal systems you already rely on, read before write and nothing that sends outside the business without a person approving it.
5. Who owns each agent, and what happens when they leave
Every shared agent needs a named owner who answers for it. When people leave, their agents do not leave with them, and an agent nobody owns keeps running on whatever it was last given. The Microsoft 365 admin center lists agents and lets an administrator reassign owners, so reassignment belongs in the leaving process.
6. What it costs and who watches
Some agents, particularly those built in Copilot Studio and those that serve people outside the business, are billed by use rather than by seat. Because that spend grows with use, decide who watches it and at what level someone is told. What Microsoft 365 Copilot costs covers the seats around it.
The controls Microsoft gives you
The agent registry in the Microsoft 365 admin center is the starting inventory, because it shows the agents in your tenant and lets an administrator block an agent across the business, limit who can activate it and reassign owners. Copilot Studio adds the data policies that control connectors. For businesses with many agents across Microsoft and other platforms, Microsoft made Agent 365, its control plane for agents, generally available in May 2026 and sells it per person.
A business of 50 to 500 people can usually govern its first agents with the registry, the data policies and a short written process. The register in your AI policy should list every shared agent with its owner, its knowledge sources and the actions it can take.
Where to start
Pick two or three agents that would help a whole team, built by people who have been shown how. Good first examples answer from your HR policies, draft first replies from your service knowledge base or help new joiners find things. Each one only reads, sends nothing outside the business and has an owner. Before opening Copilot Studio more widely, learn from those and test each against the work it is meant to save.
Where this fits
Deciding who builds and what agents may do is part of Govern. Building agents that act on your systems, and testing them before anyone relies on them, is Build, where Discover also asks whether a plain workflow would do the job better. Once they are live, keeping them owned, current and watched is Trackside.