Appearance
By Scot, IT support technician
Applies to: Microsoft 365 Copilot, Microsoft Teams, SharePoint
Article Type: Informational
Last Updated: 2026-08-09
An agent is Copilot pointed at one particular job rather than at everything. This guide covers what that means in practice, where you run into an agent during a normal working day, what it can and cannot see, who decides whether your organisation has any, and where Microsoft's own naming has not settled down. Everything below is drawn from Microsoft's current published documentation.
In this guide
An agent is Copilot narrowed to a specific job, carrying its own instructions and its own set of knowledge.
Microsoft describes agents as software that uses AI to carry out business processes, working alongside a person or on their behalf, and spanning a wide range — from a simple ask-a-question-get-an-answer helper up to one that runs on its own without being prompted.
In Microsoft's own breakdown, an agent is built from two things:
A useful mental picture: general Copilot is a colleague who has read everything you have access to. An agent is a colleague who has been given one folder, one job description, and sometimes a set of keys.
Three things separate an agent from a normal Copilot prompt, and the third is the one that changes how you should treat it.
Its scope is fixed before you arrive. Whoever built the agent chose what it reads. An agent built for a site of HR policies answers from those policies and does not wander into your inbox looking for a better answer.
Its instructions are already written. You are not starting from a blank prompt. The agent has been told what tone to use, what format to answer in, and what to do when it does not know — which is why a well-built agent gives more consistent answers than the same question typed into general Copilot ten different ways.
Some of them do things, not only say things. An agent with actions configured can write to another system. That moves it from "a draft I will check" to "something that happened", and it is the reason to read what an agent proposes before you approve it.
Microsoft's documentation puts agents in four places, and most people meet their first one without going looking for it.
One Teams behaviour is worth recognising when you see it. If an agent's answer in a group chat draws on sources that not everyone in that chat can reach, the person who asked sees the response privately first, with Approve and Reject — so a document only you can open does not get read out to the room.
This is the part most worth understanding, and the short version is reassuring.
Microsoft's documentation for SharePoint agents states it directly: agents reach your organisation's data the same way Copilot does elsewhere in Microsoft 365, answering each person according to that person's own access permissions. Microsoft's Teams documentation says the same thing from the other direction — a Copilot response in a chat is grounded in the data of whoever typed the prompt.
The practical consequences:
Understanding What Copilot Can See — Privacy and Data Access covers the full permission picture, including what happens to your prompts and what your IT team can audit.
There is one documented limit worth knowing, because it looks like a bug. Microsoft states that in Teams group chats and channels, agents built in Copilot Studio cannot use knowledge sources that require each user to sign in — SharePoint among them — and that those agents are supported in one-to-one chats only. Microsoft describes this as deliberate, to stop data reaching people in the group who should not see it. So an agent that answers well in a private chat and goes vague in a channel is behaving as designed.
Agents come from three directions, and every one of them passes through somebody's decision.
Microsoft and its partners publish prebuilt agents. An admin chooses which of those to make available, and to whom — everyone, specific groups, or nobody.
People inside your organisation build them. Microsoft offers a low-code Agent Builder inside Copilot itself for quick, in-context creation, and Copilot Studio as the fuller tool for anything more involved. Someone with a Copilot licence who can add files to a SharePoint site can create an agent for that site.
Developers build the complex ones, using Microsoft's toolkits, and connect agents that were built on other platforms entirely.
The step that catches people out is approval. An agent someone builds internally is not visible to the rest of the organisation until an admin reviews and publishes it through the Microsoft 365 admin center. Until then it exists only for its maker and anyone they shared it with directly.
So an empty agent list is a normal state of affairs. It means nobody has built one yet, or the ones built are still awaiting approval, or your organisation has chosen not to turn them on. None of those is a fault on your machine, and none of them is fixed by reinstalling anything.
What you can do with agents depends on which Copilot your organisation pays for, and Microsoft draws the line in an unexpected place.
With a Microsoft 365 Copilot licence assigned to you, agents come as part of it. Microsoft states that no separate action is needed to enable them for licensed users.
Without that licence, what most people have is Microsoft 365 Copilot Chat — the version included with eligible Microsoft 365 business plans when you sign in with your work account. There the line is drawn by what an agent reads, not by what it cost to build:
Translated for someone at a desk: an agent that answers from your employer's own files costs your employer money every time you use it, and one that answers from public web pages and its own instructions does not. That is usually the real reason some agents appear for you and others do not, and it is a budget decision rather than a technical one.
Microsoft's vocabulary in this area has changed more than once and is still not consistent between its own pages. Recognising that saves you a lot of fruitless searching.
The chat product itself has been renamed twice already: it launched as Bing Chat Enterprise, became Copilot, and is now documented as Microsoft 365 Copilot Chat. Support articles and forum answers written at any point in that chain are still online and still use the name that was current when they were written.
Around agents there are now at least two parallel vocabularies. Microsoft's builder-facing documentation splits agents into declarative agents and custom engine agents, a distinction about which AI engine runs underneath. Its user-facing documentation does not use either term, and talks instead about agents in the Agent Store, SharePoint agents, and agents you mention in Teams. Both descriptions are current, and they are describing the same things.
The clearest illustration sits inside a single Microsoft sentence: an internally built agent that an admin approves appears in the Built for your org section of the Teams app store and in the Built by your org section of the Microsoft 365 Agent Store. Same agents, same approval, two different names, one page.
The working advice: search for what the agent does rather than for the category name, and when a colleague's instructions do not match your screen, check the date on what they sent you before assuming you have a different problem.
Microsoft's own guidance warns that AI answers are not always accurate, because they are produced from patterns and probabilities in language rather than from a lookup, and asks you to verify before acting. Four failure shapes are worth watching for specifically.
Confident answers from stale material. An agent scoped to a SharePoint site answers from what is on that site — including the 2023 version of a policy nobody got round to deleting. The agent has no way to know which of two documents is the one people actually follow.
Silence about what it could not see. An agent does not announce that it skipped a folder you have no permission to open. Its answer is complete as far as its own view goes, and it reads exactly like a complete answer.
Actions that go further than you meant. An agent that can write to another system is doing something you would otherwise have done by hand and checked. Read what it proposes before approving it, particularly anything that sends, assigns, or changes a record.
A capability gap between chats and channels. The group-chat limitation in section 4 makes some agents genuinely less able in a channel than in a private chat. That is not the agent failing; it is a boundary Microsoft put there on purpose.
The habit that covers most of this: open the citations. A good agent links to what it drew on, and one click on a source tells you more about whether to trust an answer than any amount of rereading the answer itself. How to Write Effective AI Prompts for Office Work covers how to ask in a way that makes those sources quicker to check.
This area changes month to month, so treat every specific here as needing a local check.
This guide reflects Microsoft's published documentation as of August 2026. The concepts — an agent is scoped Copilot, it inherits your permissions, someone has to approve it — have held steady across the changes. The surfaces, the labels and the billing arrangements have not, and Microsoft is still shipping changes to all three.
Four things only your own IT team can answer, and none of them is written down anywhere public:
Ask them in those terms. Something like: "Do we have Copilot agents enabled? I'd like to know whether I'm on a Microsoft 365 Copilot licence or Copilot Chat, and which agents have been approved for us to use." That question is specific enough to get a real answer rather than a link.
INFO
Nothing in this guide is a setting you can change yourself. Every gate described here — licences, billing, approvals, which agents are deployed to whom — sits with whoever administers Microsoft 365 where you work. Reinstalling Teams or Office does not create an agent that was never approved.
| Symptom / Error | Potential Cause | Solution |
|---|---|---|
| No Agents section in the Microsoft 365 Copilot app | Agents not enabled, or no Copilot licence assigned | Ask IT whether Copilot agents are enabled for your organisation, and include which app you are looking in and whether you hold a Microsoft 365 Copilot licence. |
| The Agent Store is empty apart from Microsoft's own agents | Nothing built internally has been approved yet | This is normal. Ask IT which agents have been approved for your team and whether any are waiting for review in the admin center. |
| A colleague uses an agent that does not appear for you | The agent is deployed to specific users or groups | Ask the colleague which agent it is, then ask IT to add you to the group it is deployed to, naming the agent and what you need it for. |
| An agent answers well in a private chat and vaguely in a channel | Documented limit on knowledge sources needing per-user sign-in | Expected behaviour, not a fault. Ask the agent in a one-to-one chat when you need it to read SharePoint content. |
| An agent says it has no information on something you can see yourself | The material sits outside the agent's configured scope | The agent reads only what it was pointed at. Ask whoever built it to add that site or folder to its knowledge, naming the exact location. |
| An agent's answer contradicts a document you trust | Stale content in the agent's knowledge source | Open the citation on the answer. If it points at an outdated file, tell the owner of that site — the agent will keep quoting it until the old copy is archived. |
| An agent that reads company files stops working mid-month | Metered consumption capacity used up | Ask IT to check the pay-as-you-go capacity for Copilot agents, and tell them the date it stopped and the name of the agent. |
IT Tip Tuesday
Get a short, actionable IT tip in your inbox every week.
Nothing loads from beehiiv until you click. The form sets beehiiv's own cookies once it does.
