Graham Hosking: Just like we instruct Agents to do things today skills actually define a... It could be a process, it could be a thing to think about that's gonna be repeatable. So for example, like skills in SharePoint, if you think about that, you're going to put specific topics of information into a SharePoint library or a folder.

Wouldn't it be cool to have a skill to say, well, when that type of information goes into that folder, a process automatically gets kicked off for that. So going back to, like, RFPs, you know, you could put an RFP in a specific folder, and it kicks off whether that organization is fit to, to sell those types of services.

Is that something they do today? Is it something, a product that they have to think about and put in the Roadmap? Like, these are, are business processes or skills that can be defined to anything.

Tom Arbuthnot: Welcome back to the show. This week we are talking about real business Agentic workflows. Graham is excellent at decoding what the different Agentic options are in the Microsoft stack, and he talked us through an RFP use case that was a bit too much for declarative Agents. Did he need to move to pro code, or actually, as it turns out, could he do a lot of that with skills and the new Cowork in Copilot Cowork capabilities?

So really great conversation. He helps decode what the different options are and where they fit. We talk a lot about skills and where they fit in the ecosystem as well, and even a little bit of what might be coming next in terms of Agentic automation. Many thanks to Graham for jumping on the show. Great to have him on again.

And also many thanks to Barco, who are the sponsor of this podcast. Really appreciate all their support of the community. On with the show. Hey, everybody. Welcome back to the show. Really excited to have another conversation with Graham. It's been too long, and I've seen him pushing loads of LinkedIn videos that have been showing really exciting cutting-edge stuff.

So today, Graham is gonna hopefully decode skills and some of the new capabilities in the Copilot suite and talk about some customer use cases that he's been working on that are really interesting. Welcome back, Graham. Nice to see you again.

Graham Hosking: Good to see you as, as well. It's been a long, long time, I think, since we've done anything together.

I think we met in person, oh, a couple of years ago now in London doing one of the Community Days. Has it been that

Tom Arbuthnot: Long? God, that's

Graham Hosking: Bad. Y- yeah, I know. Well, we're all busy, right? We're all busy.

Tom Arbuthnot: Yeah. Yeah. No, good to see you. So remind people what you're doing now 'cause your, your role has definitely changed since we last spoke on the pod.

Graham Hosking: Yeah. So I'm looking after all of our software developers Most are software developers, what we call our SVs across Europe. So I'm one of the Solution Engineers aligned to all of them. I say the the one and only. There's only one of me that covers around kind of 53, 54 different customers around Europe.

So you can find me on planes these days quite a lot in different countries, which is nice, by the way. I quite like that. Also around- Yeah, it's good to get home today. It is, yeah. It's nice to leave the garden office.

Tom Arbuthnot: An an interesting role because it's not y- different to what you've done in the past, and I guess it's a lot of we have a pla- platform and a p- portfolio capabilities, we have Foundry, we have this

Like, how do you bring that power into your solutions? Is that part of the role?

Graham Hosking: Yeah, it's a bit of that. So the reason why I quite like talking to my customers now is it's kind of two-sided, as in we're trying to understand what they're doing as a business to make them understand what AI technologies are out there but also how they can use the Microsoft platforms and the toolings to, to provide services to their customers as well.

So I quite like that bit of the role because you're talking to them more around kind of the strategic part but in some cases going deep technical as well. So I like that kind of balance and variety. But- Yeah,

Tom Arbuthnot: But by nature it's probably a more technical audience 'cause they're building, right?

Yeah. It is. So they're probably looking at options and platforms and capabilities in a different way. But- Of course... I know from some of these businesses they, they have the business part of the business that are just going on the AI journey like any other business. So just because they have developers and a product doesn't mean they don't have the same challenges and opportunities in terms of AI adoption on the business side.

Graham Hosking: Yeah, you're right. It's... I think everybody is trying to get up to speed with how they oper- operize, authorize, authorization. Starting yet. It's a new term

Tom Arbuthnot: We've just come

Graham Hosking: Up with It is a new term. Yeah. How they start getting used to having AI do certain workloads and tasks and just workflows of things that they do today.

But really for them understanding that there's no silver bullet, that AI can just go away and fix things for them. It's not magic. So when we do things like hackathons, for example, I think people realize very quickly that it's a little bit harder than what they think because they, they go into things like ChatGPT or, or Copilot, for example, and they tell it to go away and do things, and it just does it.

Like you tell it, "Oh, I... Of all of that information that you've just found for me, can you put that into a Word document or create me a- Yep... PowerPoint slide?" So they, they're starting to use these things in a personal sense as well. But as in when they're thinking about Agents, for example- They start to realize that, like generative AI is just the brain part.

And I always give the car analogy, right? If you don't know how to drive, like you can't just jump in a car and then use the steering wheel and drive it down the road. Like you have to be taught, and that's part of the instructions. So you're, you're pretending that these are a new coworkers that you're trying to train up to do a particular task.

But again, just like driving a car, just because you've been taught, you may not have the car, right? But you need to be able to steer it, and you need to have wheels, and you need to have an engine, and that's where we talk about like knowledge and tools and deterministic sets of tooling to be able to do the actual job.

AI doesn't do the job. It decides what tools are gonna be right for that job and then execute it. So I think it's- I

Tom Arbuthnot: Like that ana- I like that analogy. That's quite good. And I can, I, I can, I can stretch it to like, it doesn't matter how good your engine is, i.e. You're, you know, you're on the latest, you're on 5.5 or you're on 4.7 or like you're, you're on the latest Frontier model, but if your tires are shot, you're not going anywhere.

So like you, you- No... Unleashing that power is, is more than just the engine.

Graham Hosking: It is. Yeah. It, it's definitely like that. So going back to the hackathons, for example, I usually run two or three-day hackathons, and you can see that when they've got all of these wonderful ideas, they're like really excited. When they start building these complex Agents, which they believe is like easy, then they go, "Like how do we connect to these systems now?"

And we go, "Well, you can connect to them. What kind of systems do you wanna bring in? What kind of data?" And they start-... To slowly get to grips about like what Agents actually means, and that, that-

Tom Arbuthnot: Well what, what is your definition? I'm putting you on the spot because it's an industry-... Term that is stretched from like Agentic in Word, i.e.

It does things for me. Like you've got, you know, OpenAI's and Anthropic's definitions. Microsoft have a plethora of things they call Agents or Agentic. How, how would you talk about it?

Graham Hosking: Well, it depends. It goes back to the business use case and deciding whether you pick one set of tooling over another, right?

So within the Microsoft suite, we talk about like Agent Builder, really easy declarative Agents-... That have some of that tooling built into them. If not, it... They're just sliders, right? We're making life easy for people out in the business to build those kinds of Agents. But if they wanna go fully the other way and build really complex Agents, like I was working with a, a customer the last couple of weeks, and we started the conversations about like Agent Builder and how you build no-code, low-code Agents in Copilot Studio, and they're like, "Great, we, we understand where we need to be now.

We need to be in Studio." And then a couple of days later, they sent me this very long detailed document about their RFP process they wanted to do. And I was like, "Well, you could. You could build it in a UI in a no or low-code fashion."

Graham Hosking: But it would take you ages. And the kind of things that they wanted to do from an Agentics point of view is they wanted recursive looping.

Now, because, like Copilot Studio is a UI and there are some, some limits to what that can do, the benefit is within the Microsoft stack is we can use things like Microsoft Foundry, and we can build Agents over there that perform more of these orchestration and looping that they want to do, and they combine the two technologies together.

So I designed that for them, and we were talking about this before we went, went live, so just some repetition for a second. But when I sat down and looked at that process, I was like, "Actually, we could do this." It's gonna take maybe seven or eight different Agents. They all call each other as and when they need to.

And I've done YouTube channels YouTube videos on this as well in the past, and the process is really complex 'cause you,

Tom Arbuthnot: Just like the car analogy- You're getting, getting to the developer end of that scale. You've gone from- Right... Declarative is basically I can get Copilot to write the Agent for me more or less to, like, Studio, okay, now I'm moving things around, I'm thinking through logic to, like, a more proper Azure-like kind of architecture.

Graham Hosking: Yes, exactly. So I'm finding it easier now to do things in pro code because you have that extra flexibility and all the Agent frameworks and all of the semantics that you need within that. And,

Tom Arbuthnot: - Yeah, and you've got things like GitHub Copilot, like, helping you build- Oh, yeah... On Azure. Absolutely. It's unblocked.

There's a different podcast for a different day about no-code, low-code, pro code, like those lines are, are radically blurring. I'm doing loads of stuff. Yeah. Like we have proper developers for our product stuff, but I'm able to dabble now on, like, stuff for the community and workflows and stuff that I just wouldn't have been able to do before.

Yeah. But I can do them- Totally... On Azure because, you know, the GitHub Copilot and Claude and Anthropic can actually help me build.

Graham Hosking: Same here. I... There's things I would never dream about going to try 'cause it would just take too long. So I, I'm engrossed it. My, my weekends and evenings are, are usually doing that and, and learning in the process as well.

So I've got a really cool video I'm, I'm halfway through recording at the moment around the thought process and the Agentic harnesses of what we call digital workers, these fully autonomous Agents that'll work in the future, and that's basically based off the, the Agent 365, like the observability piece that we talk about-

Around identity and security compliance. Like, great, you know, you can understand the chain of thought, but these fully autonomous Agents aren't something that we would interact with, right? They just go off and do their job. And if they need us, they'll come and call us, or, you know tap us on the shoulder and say, "Look- Yeah

I need some help with this now." But the thought process, I think, intrigued me, and that was based off some things I saw in the AI world with using things like Obsidian and Claude Code, like building a second brain. Yeah. And actually, as a human, that'd be awesome as a second brain, but why don't these Agents have their own brain, and we're able to see their thought process as well?

So some things some teasers there that I'm working on.

Tom Arbuthnot: Nice. I look out for that. We'll we'll cross link it when it comes up.

Graham Hosking: Thank you very much. Yeah. There's a good paper you

Tom Arbuthnot: Should read as well. So sorry, I took your, I took, I took your track. So you were buil- You did. The customer was building. You wanted to go...

You were thinking you'd have to go to the kind of Logic Apps pro code end of the scale.

Graham Hosking: Yeah, I did. I, I started building that just to take us back. So there was gonna be that, like, six or seven or eight Agents to do this RFP process, and inside of Microsoft, we just got access to, like, Copilot Cowork that was then gonna be in Frontier for everyone else.

I was like, "I wonder if we can use Cowork, that Agentic harness, and build up a series of-

Tom Arbuthnot: Let's, let's define Cowork for people 'cause that is... Like, we've now internalized it- I've never... But, like, like we've had Bas on the podcast as well. He took through it through very technically. But how would you define Cowork a- and what, how's it different to Copilot?

Graham Hosking: Yeah. Copilot, you can think of it as being able to be, have a single stream of information, right? So you're gonna give it a prompt, and it's gonna go away and give you the answer. Now, there are models built into normal M365 Copilot, where you can choose the model, so it can think deeper and harder about the process.

Or we can call on other Agents that go away and do things like deep research or deep analysis for us. So you're using a combination of Copilot and Agents to do specific individual tasks or single-threading tasks. Copilot Cowork takes that to the next level. So if you wanted it to go away and look at your emails or your calendar and say, "Look, what have I got planned next week?

I want you to go through all of my meetings, cull what I don't really need to go to, prioritize all of my emails, or start to section those out into other folders." So that's more of an Agentic piece, where it can go in and understand all of the individual tasks that we would do as a human being to say, "Well-

Tom Arbuthnot: Yeah

Graham Hosking: Is this important? Do I move this or not? Do I delete this or decline it or not?" So we're

Tom Arbuthnot: Giving it multiple And it can, it can, it can do a lot more. Like, like it's, it's, it can delete emails, it can move appointments, it can send emails. Like, a lot of the things that if you buffer it up against M365 Copilot, you'd be like, "Now send this to Lyndsay."

It'll be like, "Well, here's the copy for you to send it to Lyndsay." It's like, no. Like I want this, I want like, like the, this magic assistant. Cowork is much more powerful at doing, I think.

Graham Hosking: It is. It's super powerful. Yeah. And it, it's interesting because when you go into Cowork, initially you see like four different examples about like organizing your day, and you're like, "What is the true difference between normal Copilot now and this?"

But it's not until you start working with it and understanding, like, all of the multi-threaded tasks that it can do. So you could say, "Well, I need to get prepared for my week again, but start drafting these emails back to customers," but they wanna understand where the resources are on the Microsoft Learn site.

You can then say, "Well, I want you to go off and find some research on this particular product, and then give me the highlights of that." So you can give a prompt that has all of those multi- Yeah... Tasks in it, and enrich that with further context. So you can give it

Tom Arbuthnot: Input. And, and you can schedule as well, and, and it's running in the Cloud.

So the, the obvious comparison- It's-... Is Anthropic's Cowork, which is running on the desktop doing things on your files, whereas this is Docker containerized secure in the Microsoft Cloud, respecting all of our, you know, labeling and purview and access and all that good stuff. But off running, doing things, as you said.

So like actually it can do my... I use it for like daily briefing, and it's like, "Go check all my emails out, check out what's going on, give me a summary of what I really should be focusing on. Like, and, and tell me what is probably okay to delete," and then I can just come in and be like, "Yeah, delete those.

Like, that's fine."

Graham Hosking: Yeah, exactly. So things I'm using it for is because I've got so many customers I have to look after, as I forget what I'm doing, you know. It... They're very, very similar conversations sometimes, and you think about what the actions were that you need to do next.

Tom Arbuthnot: Yeah.

Graham Hosking: So I ask Cowork, "Well, can you build me a single page, a HTML page?

Go and gather all of that intelligence from my emails, my meetings, my transcripts, my recordings, my OneNotes-" And every morning- But you

Tom Arbuthnot: Don't, and this is the beauty, you don't even have to say where. You're just like, 'Cause it has- No,

Graham Hosking: Just go and

Tom Arbuthnot: Get it... Access to Work IQ is the kind of name for it, but all your data.

Like, I don't know whether- Yes... I had that thread in an email, or a Teams chat, or a meeting, but you've got the transcripts to the meetings, you've got the emails, and you've got the chats. Like providing you're putting transcription on all the time, it's just getting better and better at understanding what's going on.

Graham Hosking: It is, yeah. So that, we, we coin it as Microsoft IQ. So it's a combination of Work IQ around the M365 world, like your emails and calendars, but also things like Fabric IQ for data Agents, so looking inside your business data, and then Foundry IQ, which is a combination of different data sets it's recursively looking through.

So bringing all of that power into things like Copilot and Cowork just means that it doesn't matter anymore where you're having to remember where that information is. It'll just go off and find that for you. Yeah. But you can ask recursively then the Cowork to just go away and do all of those tasks, combined all of that information for you.

And in my case, I was building that single dashboard that updates on a schedule. You just say, "Can you schedule this every day at 5:00 for me?" Yeah. And it'll just build it for you, and remember that needs to be updated every day. So I've got one dashboard now with 60... No, 50, 54 different customers with what I've done with them at a high level, and then what's next.

Just makes life so much easier. So going back to the RFP- Yeah,

Tom Arbuthnot: Yeah, yeah... Story again. Sorry, I'm taking the time just so we can go over here. And for those listening, that's now in Frontier. When Graham was first looking at it, it was internal, but now it is in Frontier for- It is... External customers. So you, if you're in Frontier, you can have access to Cowork now.

Graham Hosking: Exactly. You have to be enabled in the Frontier program of course. About, as and when that's generally available, then that'll be another Agent that you can use just like you do with Researcher and Analyst today.

Tom Arbuthnot: Yeah.

Graham Hosking: Yeah, so the, the RFP piece I sat down on the weekend, had a conversation with Cowork about what the RFP process was, 'cause it was really complex, and we built out I think it was four or five different

Tom Arbuthnot: Skills.

By complex how? What, give us some feel of what were they... Like obviously you can't share the specific details, but like what types of things were they doing that made it complicated?

Graham Hosking: Yeah. So they wanted three different processes to loop. So we would go through and look at, look up the answers from a third-party system, being able to go back to the next question, and maybe have a gated process that a human would need to review.

So there was all these kind of gates where a human in the loop was required, and also- Right... Third-party systems that needed to come in as well. So you may have saw the news in the last couple of days that plugins are now available-

Tom Arbuthnot: Yeah...

Graham Hosking: Into Cowork. So we're extending that. That's huge on that.

That's

Tom Arbuthnot: Really exciting.

Graham Hosking: Yes, massive. So that's the kind of first step. The next step is gonna be custom plugins and MCP servers as well. So that'll be coming very, very soon. So MCP servers can be hosted anywhere, right? So whether- Yeah... That's using the Microsoft IQ or using your line of business application, like a lot of my customers use Salesforce, so they have their own host MCP server that-

Tom Arbuthnot: Yeah

They can use. The, the industry is kind of converging on MCP, and Anthropic obviously came up with that. OpenAI adopted it very quickly. Microsoft adopted it very quickly. I was over at Microsoft a month ago now. Every other product team was showing off their new MCP server. Like, it's been a big adoption internally.

Absolutely. And it's, it's great for this kind of use case. You say you're using ServiceNow or you're using Salesforce, like you need that data to come into the Agentic workflows.

Graham Hosking: Yeah, you're right. So as soon as that comes in, Cowork, I think, is gonna be the, the tool to do very complex tasks. And that's another common, common thing, actually, is you wouldn't just naturally just go into Cowork because it does everything.

Like, if you want a really quick answer, then you're gonna use Cowork... Sorry, Copilot. If you have a work process to go away and do, then you're gonna use Cowork- Yeah... To go away and figure all of that out for you.

Tom Arbuthnot: Yeah. Like, it's gonna- And the naming's not bad here, actually. Like, it's work, like, it's work, not query, isn't it?

Like, it's go do a thing or build a thing or have a thing run for me. Like, it's the, a, a work asset or a work process, I guess.

Graham Hosking: Yeah. And you can see it working through the process as well. So like the RFP process, whi- which was quite complicated, those skills that it started to build with me wasn't something I had to go in and write and then put into my OneDrive for it to pick it up.

It was literally a conversation about that complex process and deciding what skills need to be written to determine that process, and how it was gonna hand over from one skill to another. But similar to things like Claude Cowork, those skills can also have deterministic actions as well. So they can call scripts, for example.

So there's a, a few Python scripts that these skills call. You can also have assets as well and references. I'm, I'm glad

Tom Arbuthnot: You called that out about the deterministic elements. I feel like this gets lost a lot, and we've talked about this before on the pod, is like a lot of the pushback on AI is like, it's non-deterministic.

But it can build deterministic scripts and tools and processes. So like, like, like yes, it's non-deterministic, but when you're building these workflows, it's not the sexy thing to talk about, but a lot of it can be deterministic workflow with AI where you need it. It's not necessarily AI end to end. It's the AI's helping you build the workflow.

Obviously, the, the, the industry wants to talk about Agentic and magic in the marketing, but actually-... Workflow is a really important component of this.

Graham Hosking: It is, and that's why I bring up the car analogy. Like, you know, it's, it's one thing having the brain and knowing how to drive the car, but then you still need the car or the vehicle in order to take you there

Yeah. But you can, like say, have a conversation with Cowork to, to build up those deterministic skills, and if you think of like, like bids and tenders that all enterprises have to do, deal with today and, and build and respond to, like they're not small documents. These could be hundreds of pages. Like-

Putting that into a context window for an AI to reason over and, and think about would be almost impossible. Yeah. Maybe not in the future, of course, as models grow and get better, but you need those deterministic tools to break down the information so the AI can handle it and then decide on what they do next.

So

Tom Arbuthnot: Yeah. So you mentioned skills there, so let's define skills real... 'Cause this is a, this is again a thing that is, is, is hit the wider AI industry, and Microsoft have been quick to adopt. Just the week before last I was at M365 Conf, and SharePoint got skills. We've got skills in-... Cowork now as well, and I'm, again, I'm sure it will come to the, the entire Copilot stack, but what is a skill?

Graham Hosking: Well, you can think of a skill as like a, a reusable block of instructions. So just like we instruct Agents to do things today skills actually define a, it could be a process. It could be a thing to think about that's gonna be repeatable. So for example, like skills in SharePoint, if you think about that, you're going to put specific topics of information into a SharePoint library or a folder.

Wouldn't it be cool to have a skill to say, well, when that type of information goes into that folder, a process automatically gets kicked off for that. So going back to like RFPs, you know, you could put an RFP in a specific folder, and it kicks off whether that organization is fit to, to sell those types of services.

Is that something they do today? Is it something, a product that they have to think about and put in the Roadmap? Like these are, are business processes or skills that can be defined to anything But using those,

Tom Arbuthnot: We, we build a skill. So we build a, essentially a markdown file that says, "Do this, don't do that, use this, don't use that," and that becomes like a, a reusable asset that either for us or for other people in the business.

Graham Hosking: Exactly, yeah. But again, they're not necessarily static either. So processes change, right? So you could go back in, into Copilot Cowork and say, "Well, this process is slightly different now. Can you go in and update that skill and all of the assets that are associated with that?"

Tom Arbuthnot: Yeah.

Graham Hosking: And, and that's where, why instead of building the kind of seven or eight Agents that was gonna be in Foundry and Copilot Studio, it was far easier for that process to build it in skills and prove the process out.

Now, you could argue, is that gonna be Enterprise Class because you've just built a load of skills that one person's going to use? Or do you build some Enterprise Class Agents that have proper rigid processes to go through, like it end to end? But I can see what you're saying is because skills are so easier, easy to implement across a technology stack, like are skills now all we need to

Tom Arbuthnot: Build

Graham Hosking: Agents?

Tom Arbuthnot: It, it's interesting, isn't it? 'Cause it, it depends on the u- the use case, the, the people using the tool, the agility of the business. I don't think there's a straight answer, like, but if we're giving this tool for the sake of like two, three SMEs who run the process, that, that bit you said there about they can update the process is critical.

Like, if you have a developer build a thing for them, and then they use it for a month, they're immediately gonna have a bunch of feedback. Now they have to queue up for developer time, explain it, get it changed, run it again, realize, oh no, we slightly meant this. Like that, that's been the classic problem with business process is it took someone with skill.

That's always what we've been trying to crack with no-code, low-code. Like, like, like have someone almost on the business side be able to build the process, and AI just seems to have accelerated that. But your counterargument about robustness and like, like it's a, I don't know, if we're a pharmaceutical and we need this process to run, that's very different.

Like now the process has been tightly defined and, and it needs to go like this, and actually we don't want somebody messing with the skills because we all agree this is the process and it's audited and controlled.

Graham Hosking: Yeah, exactly. But that's been the, the whole purpose of orchestration though, versus like the old ways we used to do things through business automation.

Yeah. Is it was very linear. So you had to go from point one to five each time. You couldn't skip it. But with- Yeah... Generative AI and these skills, if you're partway through a process or you want to drop in a new part of your business process, it can be adaptable to that too. So there's an argument for both sides, I think, and instead of having like one Agent or an orchestrator or multiple Agents, they still have to think about everything.

But with skills, you're only calling that specific skill as and when you need to. So like when I'm in Cowork and I go, "Right, today I wanna do RFPs," or the next day I say, "Well, today I'm going to do my performance review," it's not having to go through all of those skills each time. It just picks up the one- Yeah

That it needs.

Tom Arbuthnot: Yeah. Development for the thing I'm trying to do. More efficient, more efficient for your context window 'cause it knows it has the skills, but it won't load them- Yeah... Into context. And, and I guess the other thing that's interesting is, like, yes, all our organizations have unique snowflake things we do, but a bunch of it is similar.

Like, we build decks, we do RFPs, we answer emails. So like, like, the, the whole space can keep optimizing these, these skills, and there are templates and examples where it's going to get better and better.

Graham Hosking: It is, yeah. And that's s- something I realized as well, 'cause I've been playing with skills in a, in an AI sense for quite a while, be- before Microsoft actually implemented it inside the, the M365 Copilot.

But also we, remember we've had plugins within the security world, like Security Copilot for, what, the last two years-

Tom Arbuthnot: ..

Graham Hosking: Or more. And that's the same principle. There's still a set of skills that do advanced threat hunting, for example, or start looking at phishing attacks. You're using those plugins or skills to do that specific task.

So it's nothing new, it's just being widely adopted 'cause it's easy for non-technical people to give intelligence as well.

Tom Arbuthnot: Yeah. We're at an interesting spot in the industry where we're talking about MCPs, plugins, skills, connectors, and the different platforms are using those ter- terms slightly differently, and even between products slightly differently.

Yeah. But like the, the, like and, and there's not one, as you said very eloquently, there's not necessarily one answer, but it's lots of ways to either add capabilities or box capabilities or interconnect systems. It's just a, it's a really exciting time to start to unlock that interconnect. So I'm really excited that MCP is coming to Cowork 'cause that opens up that third-party ecosystem much more.

Graham Hosking: Yeah, exactly. I think it'll be massive. But a- although we are technologists and we evangelize all of these technologies, so like we're interested in it, if you think of just normal people within the business that have to think about, you know, what Agents have I got available, or do I talk to Copilot, or like there's, there's things that Microsoft have announced as well, like our own version of OpenClaw.

You know, when do you use what tool then? 'Cause they start to- Yeah... Overlap with each other, and I think that can be confusing as well. Just like the days when we brought out like Teams, I know we did- a podcast, oh, years ago when Microsoft Teams was first announced, and we were like, "Where does this fit in?"

Yeah. So it just feels like that what tool to use when is coming back. History

Tom Arbuthnot: Repeats itself, right? Yeah, the O- the OpenClaw stuff is really, or that Agentic Claw model is really interesting 'cause that's next level again, where it's like-

Graham Hosking: It

Tom Arbuthnot: Is... Building things for you and doing things for you. So it's what, what, what to use when is gonna be one of those questions for a while.

Graham Hosking: It is, yeah. So the, the version of OpenClaw, like I'm using internally today, and just getting it to do just simple things for me, like do my expenses or fill in our internal forms, being able to explain where to go and what to do and not think about it anymore, it's awesome at doing that. So I can't wait till it's publicly available.

But it has all of the- So

Tom Arbuthnot: Where are you-... Same

Graham Hosking: Grounding

Tom Arbuthnot: As you- Where, when you're using it, where are you seeing that versus Cowork in your workflows? 'Cause there's, there's some capabilities that both can do. Is it, is it more that you're kind of dispatching tasks to the, the Claw scenario?

Graham Hosking: I can't talk too much about, like, how it works internally, but- Yeah

The, the way that I use all of the tools are available to me depends on what I'm trying to achieve. So if there's lots of information I need to go away and collect, it just makes sense to use Cowork to do that because it has that integration already into Work IQ, and I can tell it to create those outputs for me.

So an example to that is, like I might be getting ready for a presentation to a customer, but I want to understand what conversations we've had so far, and then build a PowerPoint file that wraps around that and our internal offerings that we have that we're gonna showcase. So rather than me having to think, "Well, I need all of these different assets," and pull all this together-

To build that PowerPoint presentation, Cowork just brings all of that together and builds the PowerPoint presentation for me as well, all using the, the Claude models. But with the, the Clawpilot, then, you know, that is another tool that that works slightly differently.

Tom Arbuthnot: Nice. Nice. Well, we'll talk about that more when we can, but that's...

It's, it's nice- Yeah, exactly And and, and it's, it kind of reflects, again, what's going on in the wider industry, which is we're trying different, different things, different use cases, but that Claw model since, you know, about December, has really taken off in a lot of people's understanding of like, "Oh, it's a, it's a thing to help me do things."

I, I wonder we had a lot of con- conversation at the M365 Conf with the product team around like- Your, your Agent having more of an identity and being, we talk about it as a coworker, but like actually being in Teams, in the meeting, doing things for you. It's a, a really exciting future, and I think Cowork has definitely got me really excited about like a lot of the use cases I would've put as M365 Copilot, like sort out my emails, sort out my calendar.

Like they're coming true in Cowork.

Graham Hosking: They are, yeah. Don't get me wrong, there is Agent mode now across the Microsoft apps as well. So you can go into Outlook and speak to Copilot to perform specific actions.

Tom Arbuthnot: Yeah.

Graham Hosking: Which you couldn't really do before with Outlook. But you're right. Yeah. With, with Cowork, you can just get it off and you're in one place, one single pane of glass to go away and write emails, and then you can say whether you want to just send the email it's wrote, it's wrote or automatically do that each time when you do that specific process as well.

So again, slightly different use cases depending on what you're trying to achieve. But I did see something yesterday I wanted to quickly mention, that with the new plugins in Cowork, there's also Windows 365 plugins as well, so you can do computer use with Cowork. So I haven't- Yeah... Had a chance to play with that yet.

I've hooked it up, but that's next on my agenda I think for this afternoon.

Tom Arbuthnot: Yeah. That's really exciting 'cause computer use potentially unlocks, particularly in larger enterprises where there's a bunch of things that don't yet have that API or MCP surface.

Graham Hosking: Yeah.

Tom Arbuthnot: Suddenly I can, I can drive scenarios. It's funny to me that like we've had that variance of that in, you know, a, a robotic process automation, I think, was one of the seven names it had in the Power Platform.

But like- Yeah... We've had the idea of driving computer workflows for a while in the Microsoft space, so it's nice to see that coming together.

Graham Hosking: Yeah. I think browser use and computer use are, are gonna be key to anything that we do, just because sometimes it's far easier than implementing an MCP or a, or an API.

They might be there, but you might not need that all the time, or it might be complicated to implement that for people in the business. Yeah. We,

Tom Arbuthnot: We both know this, there's a long tail of software and use cases and tools built internally in different industries that nobody's suddenly building an MCP wrapper to the like, the, you know, hotel booking system or whatever.

But actually, a, a, a computer use Agent can drive that like a user.

Graham Hosking: Yeah. And I guess you can see that in just the AI industry as well. I know that Anthropic's Claude was probably the, the first to do... Well, to be very good at the computer use element. But now if you see the advances with OpenAI and their 5.5 model- That's even better again.

So it's interesting- Yeah... To see, like, this shift of these models that are coming out-

Tom Arbuthnot: Yeah, we haven't got it in the Europe or the UK yet, which is bumming me out. Like- Yeah.... I'm waiting for, I'm waiting for them to unlock it here. But yeah, it looks like, like the reviews of 5.5 are, you know, for a long while Anthropic was the most, you know, competent at tool use things, and it feels like they're open AI team.

And I hear you know, on 6 it's gonna be yet again more, more capable around enterprise use cases and business workflow use cases, which is obviously my, my world I'm most excited about. So it's nice to have- Yes, it

Graham Hosking: Is...

Tom Arbuthnot: Options, and I guess that's one of the benefits of the Microsoft model is different models are available.

We- we're gonna pick the right one for the right use case.

Graham Hosking: Yeah, you're right. We- in the UK and the EU is off by default. So there, there is a DPA a, a legal agreement between Microsoft and Anthropic that obviously your business data won't be learned from by using those types of models. But the the sub-processor for the Anthropic models, it could be outside of your sovereign region.

So a lot of my conversations around that too, 'cause I look after Europe. But most organizations, like, have a development team using things like Claude Code, so there's already some sort of agreement or some special dispensation that allow them- Yeah... To use those models. And that's a- another, like, say, common question about turning on those models because you gain access to extra capabilities like the Word, Excel, and PowerPoint Agents that have those-

Tom Arbuthnot: And I think it's public now that coming on the Roadmap missing to, like, the ability to enable them for subsets of users, which is really good because- Yeah,

Graham Hosking: That's

Tom Arbuthnot: Right

There might be part of my organization that I can enable the Anthropic m- models for, and part because they deal with a particular contract or a particular customer- Yes... That we haven't yet agreed that that's a big unlock as well.

Graham Hosking: You're right. Yeah, it is. And that's, that's fairly recent as well. But again, you, you need those Anthropic models to use things like Cowork.

If you don't have them turned on, then that's not a- an Agent that you can use. So I'm sure that'll sort itself out eventually.

Tom Arbuthnot: Yes. Yeah, yeah, yeah. Awesome. Well, Graham, thanks so much for decoding that. And it's a really, that, I t- thanks for taking us through that RFP process 'cause it shows the kind of development to where we are now and how much we can do in, in Skills and Cowork as well.

Graham Hosking: Yeah. And that's, that was in weeks, like, my tack now has changed just within a number of weeks of going into complex Agents to understanding, well, do we need to prove the process out? Can we do that with Skills instead of doing all of that development time? So yeah, just watch this space. It's changing so quickly these days.

But yeah, thank you for having me on the show.

Tom Arbuthnot: No, of course. And where can people find you and your videos? 'Cause they're definitely worth watching.

Graham Hosking: Yeah. I've got two streams at the moment. One is predominantly on LinkedIn, and then I've also got a YouTube channel as well, so you can look up Graham Hosking.

And I talk about all things about Microsoft and Agents. So look out for that that Agent brain that I'm building.

Tom Arbuthnot: Yeah, yeah. We'll link that below when it comes out. So thanks Graham, and we'll, we'll not give it so long this time. A, we have to have a- Okay... A beer or coffee in person, and B, we'll- Love that

Get some more news break and then get on the pod again.

Graham Hosking: Awesome. Yeah. Thank you very much.