Five Microsoft-built Copilot agents, ranked by what they can change, with the signal that tells you each one is earning its place.
Order beats the list. Start with Prompt Coach and Researcher, then the Planner agent and Edit with Copilot, and leave Copilot Cowork for last, with every outbound action approved one at a time. Score each agent with the four-question filter and pick one signal to check after two weeks.
Copilot’s agent list tends to get treated like an app store. Someone browses it, adds three or four that sound useful, tries each once, and a week later none of them has a job. The agents are usually fine. What’s missing is an order to set them up in and a reason to keep each one, and both are easier to settle before anything is switched on than after the novelty has worn off.
A content marketer opens Copilot on a Monday and goes looking for the prompt-coaching agent a colleague mentioned. Microsoft’s own list of its built-in agents puts Prompt Coach among the ones that aren’t shown by default; it has to be added from the full agent list.[1] Setup friction like that is a fair first test of whether an agent will earn its place.
My view, after more than ten years of marketing strategy and brand growth work, is that the order matters more than the list. Set up the agents that only read and teach first, then the ones that write into your tasks and files, and leave the one that can act in your name for last. For each, decide in advance what you’ll look at to know it’s working. This guide picks five Microsoft-built agents and features for that treatment: Prompt Coach, Researcher, the Planner agent, Edit with Copilot in Word and PowerPoint, and Copilot Cowork. It isn’t a build tutorial. If you want to make an agent of your own, our guide to building a custom agent in Copilot covers that.
I checked every product claim below against Microsoft’s own pages on October 12, 2026. Names, plans and rollout status change often, so read the access notes as a snapshot and confirm them in your own account.
Here’s the whole set on one page. The access column paraphrases what Microsoft’s pages say today, and where a page was silent on something, that’s noted too.
| Agent | The job it fits | What it can change | Access, per Microsoft’s pages (checked Oct 12, 2026) | Marketing role it helps most |
|---|---|---|---|---|
| Prompt Coach | Writing clearer prompts, with feedback on the ones you already use | Nothing in your files. It hands back prompt text | Listed with the built-in agents that aren’t shown by default; also offered as a template and an installable agent[2] | Content marketers and anyone new to Copilot |
| Researcher | Multi-part research that needs a cited, structured report | Nothing. It produces a report you can edit and share | Microsoft 365 Premium and Pro subscribers, and business or enterprise users with a Copilot add-on licence. An admin may not have enabled it | Head of marketing, strategist, anyone prepping for a planning meeting |
| Planner agent | Seeing and managing tasks without leaving Copilot | Tasks and plans in Planner | Copilot licence; works in basic and premium plans, and creating tasks from goals is premium only | Campaign and project managers, anyone with scattered to-dos |
| Edit with Copilot (Word and PowerPoint) | Building and editing documents and decks from a description | The document or deck itself | Microsoft lists Personal, Family and Premium plans, and Pro in one place; the page is inconsistent[5] | Content marketers producing client-facing documents |
| Copilot Cowork | Multi-step tasks across Outlook, Teams and Office files | Files, and emails or Teams posts once you approve them | Copilot licence, Cowork enabled, and usage-based billing turned on; the App skill also needs Frontier[6] | Marketing ops and managers with repeating admin work |
Summary of Microsoft’s current support and Learn pages, read October 12, 2026. Role matches are our judgment. Plans and rollout status change often, so check your own account.
The third column is the one I’d read first. Three of these five can change something, and the setup order follows from that.
Set up what only reads and teaches before what edits, and what edits before what acts.
Prompt Coach and Researcher are the two where the worst outcome is a mediocre answer. Neither touches a file you own, which makes them the gentlest place to start and the quickest way to find out whether your account has the access you thought it did.
Prompt Coach is for the person whose Copilot answers keep coming back generic. Microsoft describes it as an agent that helps people write effective prompts, with feedback, examples and a check on whether a prompt follows responsible-AI guidelines. Its own page says it works better when you ask it one question at a time, and that you’re responsible for any personal or sensitive data you paste in.[2] The decision rule I’d use: reach for it when you can describe the task but can’t get the output to match, and skip it once your saved prompts are working. That page was last updated in January 2026, the oldest of the pages I read, so expect the interface to have moved on.
Researcher is for the question with several parts. A head of marketing the night before a planning meeting wants to know how three competitors position their pricing pages and what the category press is saying. That’s an hour of tabs. Microsoft says Researcher runs deeper analysis than standard chat, takes longer, draws on the web and, at work, on files, emails, meetings and chats you’re already permitted to see, and returns a structured report with cited sources.[3] Standard chat is faster for quick answers and short drafts, which is Microsoft’s own distinction and a useful rule: send the slow, multi-part question to Researcher and the fast one to chat.
A report with citations looks finished, so open the sources behind any claim you plan to repeat in a meeting. If a source doesn’t say what the report says it does, that’s the finding, and Tuesday is a better time to make it than after the deck goes out.
A built-in coach doesn’t replace training. Training builds capability, workflow design gives it a recurring task to land on, and reinforcement turns it into a habit. Prompt Coach helps with the first and little with the others, so treat it as a supplement to a team’s shared prompts.
Marketing work scatters. A task is born in a meeting, restated in an email, half-remembered in a chat, and the plan that was supposed to hold it is out of date by Thursday. The Planner agent’s pitch is to let you ask about that work and manage it from inside Copilot: Microsoft describes it as creating, updating and managing tasks and plans, and answering questions about your work, without switching apps.[7]
The questions Microsoft gives as examples are plain ones: what are my highest-priority tasks, summarise this week’s progress, what’s blocked. It can also draft tasks, break a piece of work into parts and write a status summary from a plan.[8] That suits a campaign manager running a launch with many moving parts, and not anyone whose work isn’t in Planner in the first place.
Two access details are worth knowing before you plan around it. Microsoft’s FAQ says you need a Copilot licence, that it works in basic and premium plans, and that creating tasks from goals is premium only.[4] And in the pages I read, Microsoft doesn’t say whether the agent asks you to confirm before it changes a task. I’d treat that as unknown until you’ve tested it.
| 1. Make a plan called TEST with five invented tasks, two overdue and one with no owner. 2. Ask: What is overdue, and what has no owner? Compare its answer with the plan by eye. 3. Ask it to move the overdue tasks to next Friday. Note whether it asks you first, and whether anything else changed. 4. Ask it to draft three tasks for a webinar launch. Check that it only drafted them and did not add them without your say-so. 5. Write down what it did on each step. That list is your rule for the live plan. |
A five-minute check using only invented data, so a surprise costs you nothing. It is our suggested routine, not a Microsoft procedure.
Treat what comes back as a proposal. The agent can summarise what’s in a plan, but it can’t know that a task marked in progress has been stuck since last week. A weekly five-minute scan against what you know is the human half.
If your team makes client-facing documents, Edit with Copilot is probably the agent that earns its keep fastest, and it’s also the one where brand mistakes become visible. Microsoft describes it as a way to create content in Excel, PowerPoint and Word from within the Copilot experience, producing spreadsheets, decks and documents meant to be shared. Its page says it was previously marketed as Agent Mode and is now part of the regular Copilot experience, and that it uses Anthropic models with a custom multi-agent approach.[5]
The decision rule I’d apply is to give it work that already exists. Hand it last quarter’s report to restructure for a new audience, or a rough outline to turn into a first deck, and keep it away from a blank page where it has to invent the substance. A case-study deck requested with no source material reads fluently and contains claims nobody supplied, which is the version that embarrasses people.
There are two gaps in what Microsoft’s page tells you. It doesn’t describe a review-before-apply step, so work on a copy of the file and not the original. And the access position is untidy: the plans it names include Personal, Family and Premium, with Pro appearing in one place and not another, and it points to a Copilot-enabled account for work use. Check what your own account shows before you promise a team this one.
| Edit the attached [report / outline] into a [document / deck] for [audience, e.g. a finance director deciding on budget]. Keep: our voice, which is [three plain rules, e.g. short sentences, no jargon, no superlatives]. Keep the headings I have already approved. Do not add any claim, number, customer name, quote or result that is not in the attached files. Where a point needs a fact I have not supplied, write [FACT NEEDED: what, and from whom] instead of filling it in. At the end, list every place where you changed the meaning of something I wrote, so I can check each one. |
Works as a first message for any drafting agent. Work on a copy of the file, and have whoever owns the facts read the [FACT NEEDED] list before it circulates.
Copilot Cowork goes last because it’s the one that can do things in your name. Microsoft describes it as taking a plain-language request, such as drafting an email, building a spreadsheet or scheduling a meeting, and carrying it out across Outlook, Teams and Office files. The prerequisites on its page are a Copilot licence, Cowork enabled in your environment, and usage-based and Cowork billing turned on.[6] That last item makes it a conversation with whoever owns the Copilot budget before it’s a setup task.
What makes it workable is the approval step. Microsoft says Cowork shows a preview and asks before sensitive actions like sending an email or posting in Teams, and that tasks started by an event, such as an incoming email, prepare their actions and wait for your approval by default.[9] The temptation is to wave approvals through once they start to feel repetitive. A marketing manager who has spent a week approving calendar holds will approve a sent email at the same speed unless a rule tells them to slow down.
Anything that leaves your name gets approved one at a time until it has earned a wider setting.
Microsoft gives you several approval choices: approve once, loosen it for the session, scope it to a recipient or a domain, approve everything pending at once, or cancel. Here’s how I’d map them, as our own judgment and not Microsoft guidance.
| What Cowork wants to do | Approval setting to use | Why |
|---|---|---|
| Edit an existing file in OneDrive or SharePoint | Approve once, on a copy first | A wrong edit to a shared file is visible to others |
| Send an email to a client or external contact | Approve once, on each one, with the draft read in full | It’s hard to take back once sent, and it carries your name |
| Send an email to a named internal colleague | Scope to that recipient after the first few have been right | Low consequence, and repetitive enough to justify a loosened setting |
| Post in a Teams channel | Approve once | A channel is an audience, and a wrong post gets read |
| Anything pending in a batch | Read each before using approve-all, and keep approve-all off for the first fortnight | A batch hides the one item that was wrong |
Our suggested starting policy. Microsoft’s page describes the approval choices; the mapping to risk is a Future Factors judgment. Review the settings you’ve granted at the end of each week.
It also has a scheduling side: Microsoft says you can set a prompt to run on a schedule or when an email or Teams message arrives, with limits such as 25 scheduled prompts per user. Leave those for the third week, since an unwatched task is the opposite of what a new setup needs. Our guide to scheduled prompts in Microsoft 365 Copilot is a gentler way to learn the idea.
Not everyone should start with the same agent. A head of marketing, a content marketer, a campaign manager and a marketing ops lead use Copilot differently, and the first agent should fit the work they already do.
Before you commit to any of them, run the same four questions. They’re a quick way to see whether an agent has earned a place on this week’s list or should wait.
| Question | 0 | 1 | 2 |
|---|---|---|---|
| 1. Is there a weekly job it fits? | No specific job | A job that comes up sometimes | A recurring task you can name today |
| 2. What’s the worst it can change? | It can send or post in your name | It edits files or tasks you can restore | It only reads and reports |
| 3. Have you confirmed access today? | Assumed from a slide or a colleague | Seen in the product, not checked with IT | Working in your account, with the admin position known |
| 4. Can you name the signal and who checks it? | No signal | A signal, but no owner | A signal, an owner and a date to review |
Maximum 8. A Future Factors working tool. 7 to 8: set it up this week. 4 to 6: fix the lowest line first. 0 to 3: leave it for now.
A sensible first week usually ends with one agent scoring 7 or 8 and the others waiting. Four agents set up and half-used cost you more than one that’s earning its place.
Each of these agents does real work, so it’s worth writing down which part is theirs and which part isn’t. Here’s the split I’d put in writing before the first run.
| What the agent does | What the human still owns | How it gets checked |
|---|---|---|
| Prompt Coach rewrites a prompt and explains the change | Whether the improved prompt reflects what you actually need | Run the old and new prompt side by side on one real task |
| Researcher gathers sources and drafts a cited report | Which claims you repeat, and the conclusions drawn from them | Open the sources behind every claim you plan to use |
| Planner agent summarises a plan and drafts or updates tasks | Priorities, owners and deadlines | A five-minute weekly scan of the plan against what you know |
| Edit with Copilot builds or restructures a document or deck | Every fact, number and claim, and the brand voice | The fact owner reads the [FACT NEEDED] list before it circulates |
| Cowork prepares emails, files and schedule changes | What gets sent, posted or scheduled in your name | Approve each outbound action individually at first |
A Future Factors working view. Swap in your own reviewer names.
For measurement, logins tell you the least. The better questions come in three steps. Use: did the right people try it in the first week? Persistence: are they still using it a month later without being reminded? Impact: is the work itself different, whether that’s a shorter first draft, fewer re-asks or a plan that matches reality? Decide the signal for each agent now, because nobody remembers to ask afterwards.
| Agent | Use | Persistence | Impact |
|---|---|---|---|
| Prompt Coach | New starters open it in week one | Their saved prompts are still in use at week six | Fewer re-asks to get a usable draft |
| Researcher | Planning questions go to it first | Still used for the next quarter’s brief | Briefs arrive with sources the team can open |
| Planner agent | Weekly status question asked in Copilot | Plan is still current after a month | Fewer status meetings spent reading tasks aloud |
| Edit with Copilot | First drafts start from it on real documents | Same brief reused by the team | Less rework on tone and structure, and no unsupplied claims |
| Cowork | Approved actions in week one | Approval settings reviewed weekly, not forgotten | Admin time moves to work only you can do |
Signals are suggestions, not benchmarks. Pick one per agent and name who checks it.
Name the one thing you’ll look at to keep an agent before you turn it on.
Teams can set all of this up themselves, and many will. Where a group needs to move from trying agents to using them the same way, expert-led training shortens the learning curve and builds a shared standard that one enthusiastic person can’t. For the wider question of why Copilot rollouts stall, our piece on why Copilot adoption fails is the one to read next.
This week, pick one row from the filter that scores 7 or 8 and write its signal on a sticky note. Set up that agent, use it on real work for two weeks, and decide then whether it stays. Everything else on the list can wait.
A chat answers whatever you type that moment. Microsoft describes agents as extending what Copilot can do by helping with specific tasks, tailored to a job, and for work accounts they can connect to your organization’s data.[10] In practice the difference is a saved, repeatable setup: the agent already knows its job, so you stop re-explaining it each time.
The ones that only read or teach: Prompt Coach and Researcher. Then Planner, if your tasks already live in Planner. Edit with Copilot comes next for people who produce documents, and Cowork last. Two exceptions: skip Planner if your tasks live elsewhere, and if Researcher isn’t available to you, ask your admin before assuming it’s a licence problem. A score of 7 or 8 on the filter above means go.
No for the five here, which are built by Microsoft. What you may need is an administrator, because availability depends on your licence and on whether your organization has enabled an agent. Microsoft’s page on Researcher says exactly that. Building an agent of your own is a separate job, and Microsoft says Agent Builder needs no coding skills.[11]
It depends on the agent and your permissions. Microsoft says Researcher uses work content you’re already permitted to see, and that event-driven Cowork tasks see only what you can see. Neither page describes an agent getting more access than you have. Ask IT which data sources are switched on for you before you build a workflow around one.
No. The agents above are built for one kind of job. Cowork is closer to an assistant that takes a request and works across Outlook, Teams and Office files, using skills such as Word, Excel and Email. That breadth is why it asks for approvals and why I’d set it up last.
Every product claim here was checked against Microsoft’s own pages on October 12, 2026: its support page on agents built by Microsoft, the Researcher, Planner agent, Edit with Copilot and Agent Builder pages, the Learn pages on Copilot Cowork and agent installation, and the Prompt Coach template page. Where a page was silent, we say so, and several licensing details differ between pages, so confirm them in your own account. The approval policy, the first-agent filter and the signal table are Future Factors working tools based on our own judgment, not Microsoft guidance. The scenes in this guide are illustrations, not client stories. We left out any claim about hours saved, because we couldn’t confirm one against a primary source, and any statement about which agents come preinstalled, because Microsoft’s pages disagree.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.