Hi everyone. In this video, I want to

talk about Copilot Co-work, an awesome

new capability that is exposed today as

a separate agent available under

Copilot. We can see it super quick. Here

I'm in a Frontier tenant,

and I can see under my agents, I have

this nice new Co-work

capability.

So,

what exactly is this? So, if I think

about Copilot in general,

we have a number of different

capabilities. It's fantastic for

brainstorming, finding information,

accomplishing a task, summarizing. I

have obviously the interactive I can

have like a a chat experience both

directly in Copilot within all the

different experiences.

There's things like analyst

to help me as my own personal data

expert. There's things like researcher

where I can go and do that very deep

longer thinking, understanding how to

solve and get me all the information

about a certain thing. They're tuned for

different types of tasks.

And so, now what we have as another type

of agent

is

Co-work.

And this is for when I have very

complicated requests potentially going

across multiple systems,

many, many different steps required to

solve it, but for potentially a very,

very long time. Now, straight away with

the name Copilot Co-work, the question's

going to be is it just a skinned Claude

Co-work? Not at all.

This is Microsoft Copilot's own

implementation of a Co-work

functionality. I.e.,

many tasks over a long period to address

a a very complicated thing you want it

to do.

So, the way it's working is

it has its own

Co-work

agent runtime.

So, there's a secure isolated sandbox

where it does the things. This is the

orchestrator of everything it's going to

do. Now, yes,

this Co-work agent runtime obviously has

to go and talk to

a reasoning model, a large language

model. So, it's going to go and for that

reasoning

talk to a large language model. The

specific large language model

is likely going to change over time as

models evolve, new ones come out, things

improve.

Now, part of what has made Copilot

Co-work possible with this very

long-running, very complicated set of

tasks is yes, Anthropic made a big leap.

I think it was November 2025 for complex

reasoning. It's Opus 4.6 model made a

huge leap in the ability to reason for a

very long time. I think days

before it went off the rails.

So, now what you can have is models have

this agentic loop that can reason to

complete the task. It can tell me what

to do next once it's worked it out.

And so, today at time of recording,

the Copilot Co-work uses the Anthropic

model for its reasoning capabilities to

create the plan

given the context and tools that are

available and to work out what it should

do.

Now, Copilot Co-work natively

understands

and leverages

Work IQ.

And that native grounding on Work IQ, so

think about yes, all of the M365,

Dynamics 365, etc., etc. knowledge,

but then

the context, the the things it has

learned about how data relates to other

data, how people relate to data, how

people relate to people, the rhythm of

business, how work is done, the types of

activities, and then specific skills and

tools. It is grounded on all of that

information,

and it's not just using the Work IQ API

piece by piece. It is also natively

hooked into

things like SharePoint

both from a data and API to do things,

OneDrive,

but then things like Outlook,

Teams,

um things like Fabric IQ,

and Dynamics 365.

It can use third-party

app connectors, services, APIs. And so,

this is a big deal about this directly

built on top of it. Other

solutions will use things like MCP APIs,

maybe computer used to interact with

client apps to do certain things, to

complete actions.

The Copilot Co-work is natively using

the solutions. It's going to be more

consistent, more reliable for any of

those interactions.

So, I you hear an analogy sometimes.

It's built natively on all of this, so

it's got the full context rather than

that per API call sipping through a

straw.

So, you get one query responded to at a

time, so the time taken, you may not

find all of the the perfect information

you want.

Now, another big thing to understand

about Copilot Co-work,

it it is running in the cloud.

It is not

running locally on your machine,

consuming your local resources,

having

access to everything on your local

machine. It is purposefully a cloud

agent. And because it's running in the

cloud, it's therefore observable. It's

auditable. I can use Purview on what

it's doing. I get better compliance,

better manageability.

And again, because it is a cloud agent

with all of these capabilities, but it

is not talking directly to your local

machine. There's no local device access.

As it creates artifacts,

it's going to actually go and save them

into your OneDrive.

And so, the things it creates become

additional knowledge. It will get the

proper labels, and we can see this. So,

if I jump over for a second,

and I just go and look at my OneDrive

folder,

and I'll look at my documents, I'll see

Co-work.

I'll see a bunch of different session

folders for where I have done work. So,

there's my sessions, and these are the

different interactions I have had with

it. If I select one of these folders, I

see any data about inputs, I can see the

outputs, and there's a number of

different files here

based on

things and work I have had Co-work do

for me.

And we're going to come back to that.

Now,

it does have the same concepts like

skills and plugins with Anthropic's

Co-work.

So, the nice thing here is it will be

able to use those skills and plugins

from the other platform within Copilot

Co-work.

Now, one of the things I think to really

understand this

is to see it in action, to see it doing

some really long reasoning, multi-step

work, how it breaks a problem down into

various steps. So, what I have prepared

is a folder.

And now you'll understand why I'm

wearing like a superhero type t-shirt.

I have a villain incident report folder

with a series of 20 different incidents

involving super villains. Just open one

of these up.

And it talks about what happened, where

it happened, when it happened, sort of a

description,

injuries, damage, how it broke down in

terms of exactly what happened, current

status. So, there's an incident report.

There's 20 of these incident reports I

have created.

And what I want to do,

as is pretty obvious,

I'm going to get Co-work to help me

understand what actually happened across

all those incident reports.

So, I'm going to ask it for a Word

document and a PowerPoint presentation

diving into everything, looking for

certain patterns. So, let's go and look

at our Co-work. So, we're going to start

a brand new session,

and I'm going to ask it about all of

those different villain reports.

Now, I'm not going to type all this in.

I'm going to paste it. I've prepared

this prompt already.

And what we can see

is

I'm asking it. I'm saying it, "Hey,

they're in my OneDrive folder,

so create an executive overview Word doc

and a PowerPoint presentation. Analyze

it for correlations. Look for patterns

across geo-clustering, severity by

villain,

trends.

Tone should be authoritative. Top five

priority villain ranking, recommended

resources allocation by region, etc."

So, I'm going to just

tell it to stop?

Okay. Go and work

on this particular ask.

So, it's thinking.

And you can see straight away, it starts

to break down

what it thinks it should be doing, the

types of tasks it's going to create over

here a little window

so I can see everything it's going to go

and work on.

But you'll notice for a second, I still

have

a prompt. It's thinking, it's doing

things,

but it's not

offline to me. So, I'm going to actually

give it another instruction. I've

decided, actually, what would be really

cool as well

is I would like an interactive web app

that shows all this in a HTML file.

So, I'm going to queue this up.

So, now sending it that additional one.

It accepted it straight away.

So, it's still working

on the other task, but now I've added to

what I asked it to do. I can work with

it. I can interrupt it. Maybe I could

ask it to hey, go and schedule a meeting

once this is done with Bruce and Clark

to discuss all of the things you're

going to find.

So, this is just going to go off and

it's going to carry on. It's going to

think for

probably many, many minutes.

So, I'm going to cheat. I'm actually

going to stop this

because if I go to tasks,

as you would expect, I did this earlier

on today.

Now, when I look at this tasks view,

firstly, these are all the ones that

I've done before.

You can see I basically ran exactly the

same request earlier on this morning.

But I can also view it in a board view,

so the ones that are currently in

progress, ones that have completed.

One of the first things I ever tried

with it was a financial analysis, and I

could see again the full progress of

everything it did. There's the output

folder of the deliverables it created

for me,

and I could go and see all of the work

it did. It took half an hour

to do a financial sort of deep dive

report for me, massive numbers of steps

and investigations and information here.

But,

let's focus on the same thing you just

saw me ask it to do.

So,

there's all the output. So, we go back.

I did exactly the same thing.

I asked it the same hey, 20 volume

reports. I waited for 1 minute, then I

added on this idea of a self-contained

HTML file.

And

here you could see it did that same kind

of thinking about what it should do,

all of the details.

It went and retrieved the contents of

the data. It tried different ways to get

the data.

It then did various query graphs to get

information from it.

It read the results.

You created to do things.

All of the data.

Then it starts structuring what it wants

to create.

Just really working out. It went through

the various incidents, all the different

correlations. Just a massive amount of

work going on, all these different

things over a fairly long period of

time. I think it maybe took 10 15

minutes in total

to complete

what ended being, as we scroll down, so

it's creating the Word doc, creating the

presentation.

There's the Word, there's the

PowerPoint, there's the

actual application it created

until it delivered everything.

So, it delivered

my Word doc, my PowerPoint, and that

HTML file that I asked it to create.

And so, there were the outputs.

So, we can open up.

There's the Word doc.

So, it gave me that full analysis,

really nicely presented, very clear.

All the information I asked it for, the

prioritizations.

It created me the PowerPoint document.

Again, same nice formatting.

You can go through, very clearly see

all the different clusterings of

information, which is exactly what I

asked it to do.

However,

the app didn't actually work.

And so, you see there was a second

prompt. And all I did was say the HTML

map does not seem to work. Can you fix

it, please?

And it then went away,

and once again reasoned for a certain

amount of time, found some problems,

and it fixed it.

So, again, it is going through, looked

at the different issues that it may

find,

and fixed all of the problems. So, I

then had the incident map app that it

created for me.

And you can see the different options. I

can run it over a certain period of

time.

It's going to show me this various

things, and I can just hit play,

and it starts over a time period showing

me all of those incidents on a nice

little map.

The amount of damage, so there was an

increased clustering of them, until I

get all 20 incidents. I can select one

to see the detail about it.

But it's just this fantastic. And again,

I can move the little slider,

and it created

an app for me.

One prompt.

Hey, look at all those 20 different

files.

And then, oh, I added an additional ask

into it say, hey, go and create me an

app as well while it was thinking on the

first thing

to generate all of that content, the

Word, the PowerPoint, and my app. And

when the app didn't work the first time,

hey, I just asked it to fix it, and it

was done.

And again, it's written into my

OneDrive. That was actually the folder I

showed you earlier. And [snorts] if

you're curious,

I used Co-work to create the 20 incident

reports. Again, I just gave it a single

prompt.

I'm working on a Co-work demo.

I need 20 different incident reports for

DC bad guys that have dollar damage,

what happened and when. I'm going to use

it to create Word doc and PowerPoint and

a fun web app. Can you create for me?

And it went and thought about it, and

that was the only prompt I gave,

and it went and created me

the 20 documents that you see. It

noticed they had a problem on one of

them,

and it's saying, hey, there was a

transient issue you need to go and fix

and rename that.

So, I actually used Co-work for the prep

that you saw

to actually make it go and do all of

this stuff.

So,

really crazy powerful.

As long as you give it some fairly

decent instructions that can have many,

many requirements,

it goes and works it all out for you.

That is the benefit here for that longer

running, more complex, this is what

Co-work does. Now, as mentioned, the

model's going to change over time. Today

is an Anthropic model. They have a sub

process, which means they have the same

global data privacy guardrails that

Microsoft has internally for all of your

services and your data.

And that's it. I mean, I really just

wanted to show it to you for the most

part because I think seeing really makes

it click the difference with what

Co-work brings over existing kind of

per task, smaller duration interactions.

So, just tell it, hey, I need this

result, and it will go and work out how

to do it. It really is a complete game

changer compared to me doing one tiny

piece at a time.

So, I hope that helped. As always, till

next video, take care.