Bas Brekelmans: I should be able to like not just manage my calendar and my emails and, and generate documents, but also if I have a receipt and I need to file an expense report, I should just be able to paste my receipt into Cowork and then forget about it. And
Tom Arbuthnot: yeah,
Bas Brekelmans: Cowork to understand that like this is a receipt, I'm in a business environment, it probably means I should file an expense report and it should, like, fill in all the fields and maybe ask you about, you know, if it was a dinner expense, maybe it needs a guest list or something like that.  but if, if I paste in a receipt and I put in some names, I expect it to do the whole thing autonomously. So again, like I want to be lazy and I don't want to put in effort into this stuff because nobody likes filing expense reports, right? so we want to make sure that like in the systems that are that matter that you do transactions in every day.
Tom Arbuthnot: Hi, and welcome back to the show. This week we are talking Microsoft 365 Copilot Cowork. This, I think is the most exciting thing happening, certainly in the Copilot and Microsoft 365 space at the moment. A massive step change in tooling and capability. And we talked to Bas, who's one of the Lead Engineers working on it about the architecture, how it works, how it differs to Anthropics, Cowork, and also some of the features that are just around the corner.  Different model choices. Really interesting conversation. Many thanks to Bas for jumping on the show and sharing so much. And many thanks to Crestron 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 podcast.  Really excited to record this one. This is my new, new favorite tool for sure. I've been really, really excited about it. We're gonna talk about Microsoft Copilot Cowork, and I think we've got the, the best person to talk about it. Bas do you wanna just introduce yourself?
Bas Brekelmans: Yeah, thanks Tom. It's a pleasure to be here.  my name is Bas and I'm a Engineer on Copilot Cowork. I, I basically started the project, here in Microsoft. And before that I was working on Copilot Studio on the agent platform.
Tom Arbuthnot: I, I want to, obviously we get deep into Cowork, but just give us a bit of your background, because I feel like you were doing this before the rest of the world caught up with, Copilot Studio and Power Platform and, agents and all that kind of thing.
Bas Brekelmans: Yeah, sort of. I think, I mean, I joined Microsoft, late 2014. and I've been a Software Engineer, you know, the, the whole time basically. and I, I worked on different parts of like business applications and Dynamics, as well as, some parts of Power BI, and customer insights. And then, there is something called AI Builders, through low code AI tools that are integrated into Dynamics and Power Apps and so forth.  and then I switched to Power Virtual Agents, like maybe it's like, I think six years ago at this point. something like that. Been working on sort of AI and seeing the evolution of chat bots and that have now become these, you know, more autonomous, agents. So yeah, super excited.
Tom Arbuthnot: Awesome. And more, more recently you were kind of, focused on Copilot Studio pre to Cowork.
Bas Brekelmans: Yeah, that's right. I was like the Power Virtual Agents that transition into Copilot Studio, and I sort of saw, seen that LLM transition even before, I think before ChatGPT was cool, like a year before we had an. Open AI model, like a GPT 2 that we were hosting somewhere in Azure. And we had it answer some like random questions, but the model like model wasn't capable enough and it wasn't, definitely wasn't able to do grounding properly and so forth.  So it was like, it was an early preview of what was coming, but we, we didn't expect that, like I think nobody understood that if you just throw more compute at those models, that they just get better. and, and now we've learned that now we're here.
Tom Arbuthnot: Awesome. So you were the, or one of the kind of, initial instigators of, the, the Cowork tool.  So talk us through that story. 'cause obviously, traditionally Microsoft has been, heavily involved with Open AI and the Open AI models, and now, and Satya has kind of said over the last couple of years more of a multi-model approach. So this is using the, Anthropic models in, in the case of Cowork.
Bas Brekelmans: Yeah, in the, in principle, right now, Copilot, Cowork, exclusively uses, Sonnet and Opus from Anthropic. but there is, like, our strategy in general is to be able to run these workloads on the best model that's available. And so that may mean that we will change, you know, the, the provider or the technology that we're using, underhood for that model.  If we can use, if we can deliver a better user experience with that. so it's not, we're not necessarily like. Hard tied to one of the, the vendors, but we feel that at this moment, for this scenario, for Cowork, that the Anthropic models do, a bit better.
Tom Arbuthnot: And can you share, like, at, at what point did Cowork come out?  Because obviously Anthropic launched their version of Cowork earlier this year. Is it similar timeframes for you at Microsoft, or were you thinking about this earlier? Later?
Bas Brekelmans: We, we are thinking about it. Or like, even as early as October. But the models, it wasn't like the scenario wasn't quite there and the models weren't quite there.  and so like we had it a little bit, I would say, on the back burner. and then like early this year we kicked it into full gear. But like the, the product right now, like co the Copilot Cowork like system and codebase and architecture is all very new. Like it was re rebooted, I say mid to late January.  Mm-hmm. of this year. So it's been about, you know, almost, it's like three, three months, almost three months now, that we've been working on it. and yeah, it's, it is an entire new system. It's also very different from, other like Anthropic score, but also other similar solutions in that we completely run in the Cloud.  so it's, it was also quite a bit more work for us to. Actually develop it because it's a, it's a Cloud service and it's doing these very complex AI workflows and we had to build all that in a very short amount of time.
Tom Arbuthnot: Yeah. So just to level set for everybody listening in, so as we record, Cowork is in Frontier exclusively, and you have to have the Anthropic models enabled to, to use it today.  That's right.
Bas Brekelmans: That's right.
Tom Arbuthnot: So, so that I, I love that you've kind of made the comparison there. 'cause I, I get to play with all the different tools. Obviously we're a Microsoft shop, so we're a Copilot shop, but we do have licenses for everything. So I, I've got good familiarity with Anthropics Cowork and I've kind of making, I think it feels like the natural comparison, right?  It's the same name, same kind of real world skills. So the Anthropic approach is kind of. Essentially, it's kind of like a version of Claude code running in a container, doing things on your computer, in a container with your files. And then kind of, can you take us through the, that, that difference of Microsoft's version of Cowork running in the Cloud?  Yeah,
Bas Brekelmans: yeah. When we started the project, there were, I was concerned because I we're used, you know, GitHub Copilot and similar tools in inside of Microsoft. And, we're, at least me personally, it was very. I'm a very heavy user of those type of tools, like to do software development and even like work item management, triage, all kinds of similar things.
Tom Arbuthnot: Yeah,
Bas Brekelmans: and the one thing that like really annoyed me is that I have to set up a Cloud VM or a Cloud machine to have it continue working while I, you know, go to a meeting or drive home or something like that. And I felt like that's not a, not a good user experience for. Anybody who's even, it's not a good user experience for anybody.  But definitely if you're not a technical, like a Software Engineer, it's really difficult to get all that like set up. And most, most of the, it, you know, departments in conference won't even let you set up a Cloud based virtual machine or people
Tom Arbuthnot: No, I know. It's very, either it's a very developer or very leading edge.  Like you see a lot of the. That kind of open claw and now with dispatch on Anthropic, like people literally leaving machines running just to do that thing. Yeah.
Bas Brekelmans: And it's causing them to do insecure things like they set up a Mac meeting or like they wire up their phone number to, and then give the, basically give like a, you know, a sort of loose, entry point access to like the deep depth their of their machine.  Things that   just
Tom Arbuthnot: won't, just won't fly in enterprise.
Bas Brekelmans: Yeah, exactly. And it's, it's like a little bit scary too, like if you're an IT Admin and you know, like everyone's starting to do it and you want to lock down even more and like block people from using things. And we thought like we need to create, a safer version of this.  And and so like when I was thinking it was like the first, I think in the first day or like after, after a few hours, I was like, we have to do this in the Cloud. Because otherwise there's no way for us to guarantee like audit and security and data loss prevention and those type of things because it's just super hard to do that from from a local machine.  Like there are technically ways with like device management policy and so,
Tom Arbuthnot: yeah,
Bas Brekelmans: which becomes super complex
Tom Arbuthnot: and like harness wise, is it. Parts of what Anthropic have done with their harness for their models? Or is it really just Opus and Sonnet, but you guys have written the, the harness and the tooling around it.
Bas Brekelmans: Yeah, so we're, the harness is also evolving. Like we have a sort of, it's right now, has been using a, initially used been using a Cloud, harness, but we also have it, it's also able to run using a GitHub Copilot harness. We're looking at just building one that is more suitable for running in the Cloud because we, we also see like the, the harness that were built, they were, it is a similar story on the Cloud hosting where they were designed to run on somebody's machine and with all kind of assumption there.  And it's not always the most efficient thing when you're running in the Cloud. We want all the power of the model to be available and, and we want it to do the right thing and have memory and all that stuff, but also be efficient when it's running in the Cloud at a larger skim.
Tom Arbuthnot: Yeah, I guess it's interesting.  You've got a, a different concern over efficiency when you're using other people's computers that you're not paying for. Then it's like, cool. It takes a couple of GigaOm, it's got access to Terminal and Grep and all these things in a, in a virtual, essentially a virtual container. Like that's not our problem, like, like compute wise.  But in, in a model where it's running in Microsoft Cloud and potentially you're gonna have, well, more than likely you're gonna have tens of millions of users, you know, if not hundreds of millions of users. So suddenly that scaling gets really interesting.
Bas Brekelmans: Absolutely. And I think it, it's also, there's some things that are different.  So when I, and you see it come out in the architecture, there's like some tools like, you know, remote control or like dispatching tools that
Tom Arbuthnot: yeah,
Bas Brekelmans: try to replicate things to your mobile phone. But running in the Cloud you can have the phone natively steer things. And I think that's also, that's one of the critical things.  But there's some, there's some interesting things like, for example, the app approve a. A tool approval. Right. it's like, is it okay to send this email and you show like a nice preview of that email?
Tom Arbuthnot: Yeah.
Bas Brekelmans: We don't want that to expire at all or like get room. And for us, like, you know, when you're rebuilding the backend architecture, we don't want to keep.  This container live forever. If you may never decide to respond to that email prompt. So we have to also make some trade offs and like shut down the VM. But we still, when you come back to the UI, we still want you to see the, it was gonna send the email, when you send the email, when to you know, actually go and send email and then like spin of the VM again and so forth.  So there's all these trade offs that. Make it much harder to build sort of a, a large scale Cloud system. It's not only about like performance and cost, it's, it, it's more about, you know, even the semantics of how things move around when you're running in the Cloud.
Tom Arbuthnot: Yeah, yeah. So like, you mentioned like tool use there and like, I think that's one of the things that, like when I first got Cowork, I was like, I, I feel like almost this is the experience that.  The marketing story wanted to tell before, like I got in there and I could be like, create a Teams meeting between Lyndsay tomorrow at 11 bang off it goes. Like, do me a, do me a summary of my diary. Send it to me every day at 9:00 AM Bang, not like. Here it is. You can copy paste it here. It is like, like, it feels like a real big jump for me as a user, in terms of what it can do.  And it feels very slick in terms of its access to the Microsoft stack, like way slicker than I've seen previously. So how are you kind of connecting that essentially virtualized Cowork instance tool, all that tooling, is that all via Graph? How's that work?
Bas Brekelmans: Yeah, we predominantly use, Graph APIs today.  like, so the, the interesting part is like we've, I think at this point we've created a sort of proprietary, when I say AI optimization layer specific to this workload, because we use quite a bit of AI in developing, like Cowork itself.
Tom Arbuthnot: Yeah.
Bas Brekelmans: And as part of that, we, I'm always, looking at, like, I use it a lot myself, but we also get like the signals from the customer feedback and so forth.  So we understand where the system is sort of stumbling or taking more time to try and do things. And so like when we have, when there's an issue, like, you know, sometimes it's small things like email folders, and. like if you're listing email folders, you get two pages and then it didn't go and look for the second page.  And so it's missing the 11th email folder. Yeah. And then it doesn't understand how to move things. Those are the kind of things that we understand and then optimize and make sure the model does very smoothly. And the idea is that the same task, that you're doing, it becomes more efficient. And the model just does, like you said, like we, we always want the model to just understand what you want and directly run to its destination without.  Taking like, you know, the scenic route to get there. and those are things that we're constantly monitoring and optimizing and it will continue to be, become more efficient and better at, at those.
Tom Arbuthnot: and anything that's
Bas Brekelmans: used,
Tom Arbuthnot: I was surprised to see scheduled tasks more or less straight away. 'cause that was something that only very recently came to Anthropic.  I don't know if you had it first or Absolutely. Around the same time. It's like, how, is that again, because it's a Cloud architecture, you can have things running in the backend.
Bas Brekelmans: Yeah. There a lot of things that we, we can do very quickly. With the combination of it being a Cloud architecture and making use of AI while developing it and doing that in a safe way.  and so like scheduled tasks is one of them where we, you know, we're like, a lot of folks on the team are very well seasoned backend engineers that understand how to build this infrastructure quickly. and then we have a, like, we have a fairly new product so we, we can iterate, we can iterate quickly and our goal is to.  Always continue to iterate quickly, on Cowork. And I think you'll see, you know, scheduled that you'll see all kinds of features come out hopefully much faster than you normally expect from,
Tom Arbuthnot: yeah,
Bas Brekelmans: us, Tom,
Tom Arbuthnot: it is a huge analog 'cause I, I've, I've been, you know, kind of, hacking my way through some of those scenarios to make them work with Graph and with scheduled tasks.  Like you could obviously do that kind of stuff, but just to see it be slick to the point where even if I didn't understand Cowork had a scheduled task. I just use the language, like, email me this every weekday at nine o'clock and it's like, yeah, okay, I understand what you want. I understand how it's gonna get done.  Here's an example. And even like, send me the example now tune that a little bit. Like it, as a user experience. Again, it just feels like an absolute step changing. I can in really simple terms, ask what I want to be achieved and, and there it is.
Bas Brekelmans: Yeah, we wanted that. I think that will remain the goal. And we need to be careful.  As, as a team, we're always, we're very carefully monitoring this like core or system of like user, like I'm personally like, pretty lazy and impatient and my attention scan span has gone, you know, down the drain over, over the course of the last two, or say few years of, you know, first being like an a sort of technology executive and then using a lot of AI, I've lost patience for doing anything and so now my, I've come to expect it to just do stuff for me.  And if it doesn't, I'll, I'll get annoyed and I'll, I'll try to fix it and so forth. And we wanted, so our challenge is to put in more and more functionality while keeping that experience of you should be a lot like you should, everyone should be okay to be lazy and impatient and expect a lot from the system.  Just solve their problem for them.
Tom Arbuthnot: Yeah. this was all my hope for Power Automate. And it never really got there to where I, like a typical user could do it. Like it's all this power. The Graph, you know, gives you all these capabilities, but you still needed to be somewhat technical to make that stuff work.  But this is the next level. You don't even see just magic Cowork Cloud tool. Go and do the thing.
Bas Brekelmans: Yeah, for sure. Nice. That that's where we want to be. Yeah.
Tom Arbuthnot: And it seems like you've adopted skills like that seems to be coming one of the industry standards now, the kind of markdown skills type format. So I think you've shipped, is it 10 or 15 skills in the box today?  And then you've got the option to add your own skills as well.
Bas Brekelmans: And we, we support the personal skills where you can sort of have your own flavor of doing things like, you know, your, your own daily brief and what's important to you and, and so forth. But there are some built-in ones that are very useful.  Of course, document generation, like Word, PowerPoint, Excel, PDF, and being able to edit those is all done through skills. similarly for something like calendar management, it'll, it keeps a memory file and it's designed to be your sort of Executive Assistant. so it'll keep notes of what you prefer and what meanings are important and who's important and where it so is it,
Tom Arbuthnot: is it dynamically?  Updating even the inbuilt skills or is it creating a different skill based on your personal preference? The
Bas Brekelmans: personal skills are a different skill that is specific to you that you can use instead of the builtin one. Right. The builtin one though, like for example, the calendar management has its own memory file that it, where it's.  Personal notes for you for the use of calendar management. So they
Tom Arbuthnot: Oh, interesting. Cool. So even if you don't go as far as making your own skill, it's still gonna listen to your prompts and your preferences and add that to a kind of memory for that scenario.
Bas Brekelmans: Yeah, exactly. So like calendar management, I use it like twice a day and at this point I can just, like when I invoke the skill, it gives me a report of all the things that it wants to do.  Which is like, you know, accept these, decline these, and you have these four conflicts, and for them I recommend this, this option, this option. And for these two where I cannot resolve it, I know that I can ask that person to reschedule that meeting. So like it'll do all that by default. And I sometimes if it doesn't understand, I might explain, it's like, because of this reason, I want you to prioritize that.  It'll keep a note of the reason that you made the decision so that next time its recommendation can be more correct. So that, that's the kind of like personalization you get even from the built-in, skills. But then you can also have your own skills. And then we have third parties that are bringing in, their skills into the, like our store and you can acquire those and, use them when you're integrating with, you know, with other systems.
Tom Arbuthnot: Nice. So that was gonna be one of my questions is obviously like the, the MTP or or API interconnect to other platforms is a huge productivity unlock. So right now. I think in the release I'm using, there isn't any third party connectors yet, or have I just not discovered them?
Bas Brekelmans: Yeah, so for, this was one of the first things we, we were thinking about that is very important for, our product is I should be able to like not just manage my calendar and my emails and, and generate documents, but also if I have a receipt and I need to file an expense report, I should just be able to.  Paste my receipt into Cowork and then forget about it. And
Tom Arbuthnot: yeah,
Bas Brekelmans: Cowork to understand that like this is a receipt, I'm in a business environment, it probably means I should file an expense report and it should like, fill in all the fields and maybe ask you about, you know, if it was a dinner expense, maybe it needs a guest list or something like that.  but if I paste in a receipt and I put in some names, I expect it to do the whole thing autonomously. So again, like I want to be lazy and I don't want to put in effort into this stuff because nobody likes filing expense reports. Right. so we want to make sure that like in the systems that are that matter that you do transactions in every day,
Tom Arbuthnot: that Yeah, you would do your, your Jira, your ServiceNow, your, whatever it is.  Like, like obviously the first part. Yeah. I should be able to
Bas Brekelmans: do it for you. Yeah, that's right.
Tom Arbuthnot: So how are you think so in like, again, comparative tools from other vendors? It would be. MCP connectors and they do have a kind of store, but also you can load your own in Cowork again 'cause you are managing it.  Is it gonna follow that same model of like working with having an authorized connector with a partner?
Bas Brekelmans: Yeah, so, so of course like the list of partners that are available is expanding. but it's, right now we support two elements in this extensibility. There's skills and then there's MCP, servers.  The MCP servers are remote MCP servers, so it's like, you know, some of the well-known vendors will be there. but if you have your own MCP server internally that you're using for an application or if you have an API, and you can expose an MCP endpoint for it, that's also something that you can bring into Cowork
Tom Arbuthnot: and Awesome.  So if I've got a line of business app I built in-house, if I'm prepared to put an MCP wrapper around it and then make it available, I could do that.
Bas Brekelmans: Yeah, exactly. And then, so whatever people are doing in your line of business app can now be steered from, from Cowork. And also information that's being retrieved.  You can, like there's a connectivity to fabric and a lot of those will just output, like Cowork will execute queries and get data. Then can use that data to do other things. So you can then use it to, I need a spreadsheet with this, or I need a forecast on this data. You can actually, if your API is able to return some kind of CSV or JSON that has the data, Cowork will be able process it and do analysis on it and so forth.  So it doesn't, it's not necessarily restricted to. Tools that have like smaller outputs, it could even like download an entire file and that's totally, totally acceptable. Yeah, and I definitely encourage, like building MCP servers for some of the line of business application piece, it'd also be a delight for the users of the application if they have access to Cowork.  they can automate a lot of the things they would do in the, in the app.
Tom Arbuthnot: Awesome. So in, in the, in the preview window we've had so far, like I'd be really interested in what some of your use cases are, but also you've no doubt had lots of customer feedback. Like, have any use cases particularly stood out to you as being interesting or things you might not have thought of that customers have seen value in?
Bas Brekelmans: Yeah, I think I, I use it a lot to like, to a few things. Like I use it to keep track of my Teams and email and calendar, like just the basics of, because I got lots of messages on Teams. We had to, you know, pause as a podcast recording slightly because my Teams keeps going boom, boom, boom. Yeah. so just making sense of that and keeping track of that.  And the same with, with incoming emails. As well as the, tools for like calendar management, because I also get a lot of people that just drop meetings on the calendar. And I need to, yeah, make sure I'm protect enough free time. That's
Tom Arbuthnot: my, that's my killer feature. And that was again, from day one of, of, Copilot.  It was like, like, just. Make meetings, cancel meetings, control my calendar like that. I feel like that's most knowledge workers', primary pain point that and email overload and Teams overload.
Bas Brekelmans: The other thing that's, very, I found very useful is like sending status updates and reports. so I'd have a bunch of raw data about like usage and the statistics around like quality and user feedback and so forth.  And I want to generate like a nice status update that I can send to our, our leadership. And there, like, what I found actually very interesting is that. It is very good at creating HTML files and, and markdown files, but I use it to generate like a nicely formatted HTML file and I can actually email that to the audience and I get a nicely formatted email report.  And if people would ask like, well, it said sent by Cowork, so they're like, Hey, what tool did you use to generate this report? It looks really nice like. And I didn't do anything. I just make Cowork, like write nice with them. Yeah. So those are also like good use cases. and then, you know, creating PowerPoints and documents.  Say somebody asked me to send them a, like a one pager about some information they need for internal infrastructure stuff, and then I can, I can just say, Hey, Cowork, can you like make the one pager that somebody's asking me in that email and it does it. And then before I did it, it told me like. I have enough data to do like 60% of it, but I need these five other things, which I can, I then, but because we don't have connectors for everything yet, I can feed it to my GitHub Copilot, have that one, produce a report from some internal stuff.  Mm-hmm. And then paste that back into Cowork and then it did the rest. So there's like, you know, some. Some wire up that's missing for it to do in one shot, internally because there's some internal systems that we have to connect to, but like it's, that's saving me like three hours of like randomly doing, pulling together data and writing reports.  And I, I personally, I hate writing reports and status updates. And I like seeing them, but I don't like making them. Yeah. So like I, I'm, I'm very happy like that. I can just be the manager of that process and review it and say it's wrong, this wrong, and then, but not have to actually do that work.
Tom Arbuthnot: Amazing.  And there's been some more feature drops recently. Do you wanna take us through some of the, the newer things you've added?
Bas Brekelmans: yeah, I think,. A lot of it has to, like, we've, we've done a lot of infrastructure improvements as the, the usage has picked up and we've seen like where people are struggling and so forth.  So a lot of it's just like hardening and stability and making it more reliable. And, we've added, better audit logs, like richer auto logs. and then we're doing a lot of like making sure that everything that we're doing is very compliant.. So the per, like, everything shows up in Purview and there's support for like e-discovery, legal, everything that makes our IT admins very happy.
Tom Arbuthnot: Yeah. I, I feel like this is one of the crosses, you at Microsoft have to bear that some of the, you know, like, newer companies to the space can, you know, particularly the smaller ones can just fling things out there and be like, this is amazing. It's like, yeah, but like, like you have to have the enterprise user in mind all the time.
Bas Brekelmans: Yeah. Be because,. It's a serious, like, I would say it's, it is actually very important that we, we build tools where you can do like proper risk management, right? Mm-hmm.
Tom Arbuthnot: So
Bas Brekelmans: if it's, it's fun, it's fun that something can be done automatically. But if I have, if I have no visibility and control over it, then it also, it's equally scary as it is powerful.  And if I, if I can see, I can control it, I can have visibility on it, I can control like which systems it can interact with and which users have access to it. Yeah. And so forth. Like, it, it becomes a lot better. So I think that's like, probably we're still gonna be our main focus, for the, or for the coming, I would say like few weeks and months is to make sure that that is like solid and, and very robust.  But then we also have, you know, bunch of quality of life things, like the, you know, the, mobile, the extensibility with MCP servers, and then, in general improving just the quality and the flow of the AI system and making it also easier to debug these things. So if you have your own skill.  I can tell you like, Hey, you know, that skill is sort of getting in the way of this other skill, and I'm, I can help you fix it. Hey, you don't have to do anything. I can fix it for you. Yeah, I can change your instructions here and there. And then now your skills, runs more smoothly. and being able to better learn from your bare memory, from like understanding which decisions you typically make, so that next time.  I can be like, Hey, do you want to do it this way? Because that's how you normally do it, so it learns more from you and makes the whole process more efficient. so those are things that we're, we're focused on right now.
Tom Arbuthnot: How, how do you think about it from a user experience point of view? Like obviously right now customers would have to be Frontier.  We've still got the Anthropic EU data boundary stuff, so not everybody's gonna have access yet, but like when people do have access, when would they use? Copilot. And when would they use Cowork? Like do you foresee they live more in 'cause Cowork I feel like can do the same things Copilot can do and more.
Bas Brekelmans: yes, there will be some things that will announce on this, in the future. I, unfortunately I cannot speak to like where, that's
Tom Arbuthnot: fine. So the next, the next podcast Bas.
Bas Brekelmans: Yeah, the next podcast. We'll talk about it. Yeah.
Tom Arbuthnot: Yeah, yeah. Yeah. I think that's really interesting, like kind of one of the,. One of the achilles heels of the Multimodel multi-tool world is like, I, I know the, the, the business users I talk to are not following all the podcasts and all the feeds on.  Is it, you know, do I want to use Sonnet or do I wanna use Opus or do I want to use whatever it is, Five/Four. Like, I feel like more and more that needs to be for a business user paired down to like, here's the experience.
Bas Brekelmans: Yeah, exactly. And Cowork, like the default is auto. auto is what we see works best for most people.  Mm-hmm. and then we, we also have some adjustments to the, like, if you start doing deep research or certain, like we might throw in some Opus here and there to make it, higher quality depending on the tasks that you're doing. And so we try to. Like as long as you select auto. I think over time that'll be the best experience for any user in general.  Like depending, no matter
Tom Arbuthnot: what you, as a, as a as, don't get me wrong, as a power user, I like it. 'cause I could be like, oh, let's see what's better. Like I'm not, I'm not paying for the compute, so I'm like, I'll just use whatever I want to use. But as again, as it scales, I guess from a Microsoft point of view, and you've got your own.  First party models for certain use cases as well. I can see where it would be interesting to optimize auto to be like, we have a great model here that's really efficient for this and for this one, we need this model. And for this one, we need this model.
Bas Brekelmans: Yeah, that's right. and we'll, we'll continue to like, iterate and evaluate that.  And because we have, we say like, you know, the large scale we can, we can make that investment and make that determination and say like, for this, you know, for this scenario, this is best.
Tom Arbuthnot: Yeah. Awesome. Well, Bas any, anything else you want to, leave people with in the community as far as, things to try or feedback or anything else?
Bas Brekelmans: No, I would just say, be, be as lazy as possible. Basically. Stop like, this is, this is quite a transition. Like we, even our, our own, like in our own engineering team, we noticed that like just the entire way of working has changed and everyone has become like a sort of manager of the agent and Yeah.  Yeah. We've all sort of become, we've come to increase our expectations of like what the system can do. and
Tom Arbuthnot: I feel like if you've, if you've been on the Copilot journey, you kind of have to reset yourself again and retry everything that you might have tried and, and said, oh, it's not ready for that.  It's not really like it's, it's a whole, even though it's Copilot, Cowork,
Bas Brekelmans: it's just how you work. Because even in the past, like we, of course we have Copilot would ask questions and it can do some basic interactions, but. At this point, it can automate like entire tasks. Like for me, like half of the emails I get at this point, I think I, I can just throw at Cowork and it would resolve it in a way that's at like, at least this, it does.  So people are worried that sometimes the AI, it may not be perfect. The baseline is not perfection. The baseline is how you would do it and you're probably doing hundred.
Tom Arbuthnot: The baseline for a lot of us is not gonna get a reply at all. So already Exactly. Some reply is better than no reply.
Bas Brekelmans: Yeah. Yeah. And sometimes no reply is better than you know.  Well, yeah. It depends on the reply, I suppose. But I think, there is like something to be said where, because for many things, they're fairly straightforward and, like you'd spend time typing out like menial things that. You no longer have to do it. And then also navigating between tabs and systems. Go here, click there, do this, and so forth.  Like those things, making them go away. It's like, for me, it saves at least, I would say like 10 hours a day because it's, it's kind of weird because I also work a lot, but like when I, it's, it literally saves me that amount of time because of every hour there's some stuff that I'm doing with it that makes.  Like a bunch of work go away. And that's like at least every hour it makes at least an hour of work go away and I can focus on things. Yeah. And that, that's like the, the magic. And I would definitely encourage people to try and set the bar very high and don't do manual work. Like try to make the AI do it first, because you also get better at understanding it and learning it and being able to prompt it because it's very good at following your instructions.  And the better you get at writing those instructions, the smoother the experience will be. And you become basically this. AI manager. but it requires some practice to together.
Tom Arbuthnot: Yeah. Awesome. Well, Bas thanks for jumping on the pod. Really great conversation and I'm really impressed with the, just the preview of Cowork and what's the shipped so far and I'm, really excited for what come, what's coming.  So yeah, when the time's Right. Love to have you back on again and we can talk about the, the stuff you can't quite talk about yet.
Bas Brekelmans: Yeah, for sure. the team I will say like, we have a small but very passionate team and we're all very excited and we're using it every day and we love the tool as well.
Tom Arbuthnot: Awesome. Thanks a lot Bas.
Bas Brekelmans: Thank you.