AI, Actually – Episode 14: Autonomous Agents in the Enterprise and How AI is Disrupting SaaS

Pete Reilly (00:56)

And this, you know, it’s interesting if you’re watching SaaS stocks and a lot of them are delivering great results, but the stock stocks themselves are really suffering. And so we can talk a little bit about, you know, this idea of should you build your own enterprise software? Should you build your own replacement for Salesforce and so on. so with that, we’ll just jump into the first segment around.

something that has changed names, I don’t know, three times at least, the last couple of weeks, and Clawdbot open, you know, to MoltBot for a minute, and then now OpenClaw So basically it’s this, my understanding, and you guys chime in, it’s this open source AI assistant slash agent that doesn’t just suggest actions, it actually does things, right? It will send emails, it will schedule meetings and run workflows, and it behaves a little bit like, a little less like software, much more like this junior.

employee with credentials and access to everything. ⁓ first of all, I’ll sort tee it up off if you guys have better descriptions of that, that would be great. And then we can dive into man, what are some of the implications of this? Who wants to kick that?

Stew (02:02)

I think the one

Shanti Greene (02:02)

Yeah.

Stew (02:02)

thing we can agree on is that,

⁓ OpenClaw a much better name than MoltBot that’s, my main, you know, with,

Shanti Greene (02:08)

I don’t know, I already

Jim Johnson (02:09)

Thank you for that.

Shanti Greene (02:12)

had to rename my Raspberry Pi and change some domains. It was really annoying because I was like, man, was going to commit to Moltbot and now I have to change it.

Stew (02:16)

you

Pete Reilly (02:16)

I’m literally there.

Yeah. So how would you guys describe it? Did I describe it roughly correctly? Would you sort of tweak that some?

Shanti Greene (02:26)

Yeah, I think the original developer set it up to be his personal assistant and he wanted it to be able to take actions like a good personal assistant would. And it was starting to do those things. So if you think about it in those terms where you say, it’s a personal assistant, you can communicate with it. And based on that communication, it will start taking actions. And the more tools you give it access to the more different things it will be able to do for you.

Pete Reilly (02:51)

Have you installed it, Shanti?

Shanti Greene (02:53)

multiple places. Yeah. So I’ve got, I’m running it on a Raspberry Pi on my desktop. I’m running it on a cloud instance somewhere else. The setup isn’t quite as easy as you’d like it to be. Like it’s pretty easy. The developers did a great job of trying to guide you through it, but running it on a remote system instead of locally, there’s some little tweaks to how you actually redirect network traffic that can be kind of frustrating.

Pete Reilly (02:55)

Yeah.

Yeah, it resulted in a bit of a run on Mac minis. Did you get, did you, did get?

Jim Johnson (03:21)

Yeah, it was the biggest selling iMac mini

instigator in years.

Pete Reilly (03:26)

Yeah, messed up

that supply chain pretty good.

Shanti Greene (03:29)

Well, what’s interesting

is you don’t need much power. Like I’m running it on a PI 4 with 8 gigs of RAM because it doesn’t do a whole lot of processing locally unless you want to run like local transcription models, which you can, but you could run it on just about anything.

Pete Reilly (03:45)

and you put it on that box just to isolate it.

Shanti Greene (03:50)

It was the closest one to me. It was within arm’s reach.

Jim Johnson (03:52)

There

was no architecture behind it. was just whatever was convenient to shine at the moment.

Shanti Greene (03:58)

I have

Pete Reilly (03:58)

whatever was handy

Shanti Greene (03:59)

a lot of mini PCs around. So I’ve got one where I run a whole bunch of other containers, but the deployment for that one just would have taken a couple of extra steps and I didn’t have anything running on this one. So it was like burn an image and then I can install everything directly without having to worry about it.

Pete Reilly (04:02)

Thank

Right. Stew, you, did you mess with it?

Stew (04:17)

I yeah,

yeah, a little bit. mean, I think there’s a lot of interesting things to look at and, and I think it’s one of those things where it’s less about what this is today than what it reveals about the future. So I think the, you know, I think what we have today is definitely, you know, with this particular tool is definitely, I would call it a.

Shanti Greene (04:34)

Yes.

Stew (04:43)

hobbyist tool is probably ⁓ like CSOs nightmare, you know, in terms of the security implications, ⁓ the way that it’s currently implemented things along those lines. But I think it reveals ⁓ some very important things about what’s coming in the world of AI and has given people, I think, opportunity to see that ⁓ with a little more clarity, you know, than ⁓

Shanti Greene (04:49)

Yup.

Mm-hmm.

Stew (05:12)

that maybe we’ve had to date. I think there’s three main things I picked up out of it, looking at it from this standpoint, ⁓ in terms of what ⁓ working with an agent is like differently with this type of thing that is causing the freak out. And I think there’s three key things that, and I think we’ll see this a lot now migrating into enterprise design ⁓ over

Jim Johnson (05:30)

.

Stew (05:41)

the coming year. So one is ⁓ agents have. So the three things are agents that have their own identity, agents that work proactively and agents that coordinate with each other to get work done right. And in ways that were not pre-engineered. And so I think those three things are really important. And you know, I think

Pete Reilly (05:50)

Hmm.

Yep.

Stew (06:10)

The first one is around identity. I think as these agents become more and more capable, it’s important, I think, for them to have their own ⁓ security context, if that makes sense. And I think as they become more and more like a colleague that you’re working with versus a co-pilot that’s just helping you in a chat session.

I think it’s also important that we have ways of talking to and with this agent. So I think a trend we’re going to see that’s revealed by this is actually giving your agents their own identity, meaning their own potentially email context, their own Slack context, their own way. So kind of sandboxing in the universe that they have to do as well as just giving a way to refer to that person, hand that person work, do all those other sorts of things.

The second bit I think which is really interesting is just the whole structure of this thing was just to make it super proactive, right? ⁓ And so basically it’s just on, you know, scheduled jobs that wake up and are like, what can I do right now to help, you know, help the, ⁓ human? And I think that’s within an instruction set that helps it move in that way. So I think that’s, we’re just getting more and more glimpse glimpses of how that, ⁓

Pete Reilly (07:22)

Thank you.

Yeah.

Stew (07:38)

will become more and more, ⁓ I think as these agents get more capable. then the third thing, which is really interesting and probably sparked a lot of the internet conversation was just these agents, you know, they’re real glimpse of like, wait, what have we unleashed? I think started to come from this idea of, giving the agents a way to talk to each other. Right. And so, you know, for those that don’t didn’t follow it as closely, like basically they built a Reddit.

Pete Reilly (08:01)

Right.

Stew (08:08)

for the agents to talk to each other, right? So their own kind of social network for the agents. I think a lot of the activity that happened on there was probably kind of human engineered, you know, kind of like, like, doubt all of it was truly organic and what was happening and even where the bots were interacting, I think it was relatively shallow. You know, they weren’t having deep conversations with each other. Maybe they were just.

Pete Reilly (08:21)

Absolutely.

slide.

All

Stew (08:36)

posting on a thread and then moving on to the next thing. But I think it revealed some interesting patterns. And I think people that were looking forward to things really saw this and said, this is interesting. And seeing the early days of these agents having their own social network, their own way to communicate even some of the crazy screenshots that came out of those interactions, even if the human

Pete Reilly (08:41)

Okay.

Yeah.

Stew (09:06)

might have nudged it in the direction of having some of this crazy stuff. I think it reveals what’s possible, right, and where

Shanti Greene (09:07)

you

Stew (09:14)

these things could go. to me, that was like what was really interesting about this. And if we think about enterprise context, I think where this is likely to go in the future.

Pete Reilly (09:23)

Go ahead, Shanti

Shanti Greene (09:23)

Yeah. And I’ve

seen some of that in some workflows that I’ve talked to people about where they’ll take something, a tool like linear, which is nice plugins and plug that right into Claude code and have proactive agents set up that go look for new issues assigned to you. And if you see a new issue, go start working on that issue. Or I’ve got a code review agent set up now. So anytime I submit a pull request, it automatically gets picked up by the code review agent and does a AI review of that pull request.

Pete Reilly (09:38)

And.

Shanti Greene (09:50)

So getting that nice without having to do anything proactive like, hey, this has been reviewed at least once for security flaws. Okay, great. Now that’s one less step.

Pete Reilly (10:01)

If I’m just a business person, I’m not writing stuff in Claude code, and I barely use ChatGPT and sort of ⁓ know, Claude on the web to maybe ask a few questions. Explain to me in really simple terms, what’s the difference here? What’s the difference between my normal use of

I don’t know, OpenAI or Claude in a chat experience versus what we’re talking about here in now called OpenClaw. ⁓ What’s the easy version to sort of understand?

Stew (10:35)

I think it’s more that you’ve given like basically the way people are using this is you’ve given this agent its own computer effectively and you’ve given it access to most of the tools you need to do useful work on a computer, right? Email WhatsApp in many cases, you know, all sorts of tools that people could communicate and then given it a broad, you know, a broad set of instructions.

And it is ⁓ just trying to make your life more helpful. It’s, it’s like an early Jarvis, right? It’s like baby Jarvis and I.

Pete Reilly (11:11)

So we’re not taking

turns like I might in a ChatGPT or something. It sort of has a mission. It has a set of goals.

Stew (11:15)

Yeah.

Yeah, and you could send it a message.

You could send it a WhatsApp message or a Slack message or whatever you connected and just say, hey, can you schedule this thing for me? so what you can kind of do with some of these emerging tools, but I think it’s just much more set up to interact with the way you would interact with a remote assistant.

Pete Reilly (11:40)

Mm-hmm.

Shanti Greene (11:40)

Yeah,

I think part of it is that it’s listening for you so that you don’t have to tell it everything. It’s watching the social channels that you’ve given it access to. So when it sees something come across, it reads all of the things and says, ⁓ this is something I should take action on based on the directives that it was given. It will then go and see if it has the tools to take that action. So in a different tool, I’ve set up something like that in ClickUp with their super agents.

so that if I add something to a list called ideas for writing ideas, it will go check to see if somebody has already written about that topic for me. So I don’t have to go do that lookup. And then it’ll come back and let me know like, yeah, don’t bother writing about this. There’s plenty out there or else I know like there’s a gap there. Maybe it’s worth pursuing.

Jim Johnson (12:25)

To me, that’s the potential immediate intersection in the enterprise space because

Stew (12:25)

⁓

Jim Johnson (12:33)

I mean, mentioned that, Stew, you mentioned Jarvis. We all sort of continue to think about, or not all of us, but many people in the public domain continue to think about sort of the Jarvis agent that can sort of do it all and has sort of a broad-based purpose set. And I think then we see some of the other things that are coming out here, some of the concerns over sort of the security, you know, the issues around security and the issues around what actions are taken and the issues around sort

what the agents are really doing. And that messaging to someone who’s not living in AI every day is unsettling to the business community to think about agents doing work in their domain. But when you take the example that Shanti just described, which is, hey, we can set it up to proactively do work, but if we narrow it and keep it tight, and the narrow use cases continue to be where the enterprises are having success.

tightly defining the work that needs to be done, tightly defining the constraints and the ability to evaluate the results and who owns the results. That’s where the success is. To me, this is a step forward in terms of what you could do at the enterprise level, but it creates a lot of concerns when we start talking about the Jarvis model or sort of the do-it-all agent. That’s unsettling to people who are looking at this, asking the question of how can I use this in the enterprise. This is really

cool and fun ⁓ for the 36,000 people who’ve signed their agent up to go and sort chat to other agents. ⁓ But we’re already seeing sort of crazy

security issues emerging, prompt injections and bad actors are introducing sort of bad tools that could be used by the agents. It’s sort of counterproductive in terms of where we want to take the enterprise in terms of doing valuable work. It’s all going to come back together eventually, but man, we’ll this just to prove it.

Pete Reilly (14:40)

It does seem to be the model

of how agents will work and accomplish meaningful ⁓ tasks. On the one hand, on the other hand, it is scary, very frightening for any CSO or any large organization. But it almost feels like if you’re going to have an agent that’s doing these things, it’s going to work like this. How do you, but maybe Stew for you, how…

What’s your advice to, I don’t know, an executive in a larger enterprise about how to approach this? I mean, it could be ignored entirely. It could be thrown on everybody’s laptop or…

Stew (15:16)

Well, I think it’s definitely early

days on this. In fact, I was reflecting, you know, there was ⁓ in early 2023, there was a similar kind of craze that grew up from open source stuff that was they literally called it baby AGI. And it was it was basically the first time someone took

Pete Reilly (15:35)

I remember that. Yeah, yeah, yeah.

Stew (15:40)

these LLMs, and it was just a hobbyist, took these LLMs and just put them on a loop, right, in a really serious way. And you started to get some weird emergent behavior, and you could see that it was really interesting in coding situations and things like that. And it took a while, but basically that’s what became cursor, right? And then that’s what drove Claude Code and…

Pete Reilly (15:47)

you

Thanks.

Stew (16:06)

and some these other ⁓ agentic things. So I think you have to look at it as a glimpse into the future. I think for enterprises, the glimpse to the future here, think, I would say in an enterprise context, you’re gonna wanna, as Jim mentioned, you’re gonna wanna keep things much more on rails, ⁓ probably. You’re gonna want your agent coming into the office wearing a suit and all buttoned up and much…

Pete Reilly (16:24)

Mm-hmm.

Stew (16:32)

under much more tighter constraints than when they’re out and, you know, in the club partying over the weekend. ⁓ So you’re going to put this guy on a lot tighter rails, but I think it’s a, a glimpse into the way, ⁓ what I would be thinking about as a, as a CTO now ⁓ is this idea of actually onboarding non-human employees. And I think that’s what we’re going to.

sounds really weird today, but I think sometime probably this year, you’ll start to see progressive companies doing this. And what I mean by that is like they’ll, you know, we’ll have agents, we’ll give them their own email, we’ll give them their own Slack, we’ll probably tie them to a human who’s accountable to them as basically the supervisor of that agent. ⁓

Pete Reilly (17:27)

Can we start a company to do background checks on agents?

Stew (17:32)

It could, ⁓ they’ll probably

Shanti Greene (17:32)

That would be interesting.

Stew (17:33)

be, ⁓ you know, it’ll be interesting. I, and I think that’s the other interesting thing here is memory. Right. And there’s some early, we didn’t really talk about memory in this ClaudeBot thing, but, but I think there’s some early tastes of, of what memory does in terms of creating a distinctive agent, right? That’s an artifact of the context you’ve given it. And so I think that’s.

really what’s going to start to come in enterprise. And I think the original, you know, the initial employees, employees that do this will probably be ⁓ narrowly based, but you’ll interact with them the way that you do the way that you interact with ⁓ a lot of remote employees, you know, in terms of talking to them over Slack, you know, sending an email, getting responses back.

that are maybe a Word document or an Excel document or some artifact along in those ways. And then they’ll be interacting with the same systems you do, right, in terms of ERP, CRM, other type of systems, including other agents that are working on the behalf. So I think it’s that idea that we have to start wrapping our head around. This gives you, I think, some insight into what that might look like. It gives you…

Pete Reilly (18:43)

including other agents, other employees.

Shanti Greene (18:46)

Yeah.

Stew (19:00)

some insight into where that could get terrifying, right? If you don’t have the right guard rails around these things, I think it gives you, ⁓ but it’s just an opportunity, I think, to reflect deeply on this and start to ⁓ feel what this experience will be like. Because I think it’s almost certainly coming to some level, right? Whether we want it to or not, in some cases, I think this is starting to happen.

Shanti Greene (19:24)

Yeah, I think it’s gonna,

I think it’ll look like LLMs did, which is if you’re an enterprise and you’re trying to prevent your employees from using LLMs, they’re gonna use them on their phone. They’re gonna use them in non-corporate channels. If this picks up and people start using them in their lives, they’re not going to accept not using it at work as well. So figuring out what that needs to look like, how do you build systems around these in a way that are acceptable to your corporate mission?

Makes sense. Now, to some degree, your agents might follow the employee policies and playbook better than your employees, your human employees will, because they’re slightly less incentivized to go around them. Although I did have an experience back with OpenAI’s operator, back when it still called that, where it was running a long running task and told me like, this is taking a while. Do you really want me to do this? I’m like, yes, yes, I do. And then it did like one more round of it. was like,

is no, it’s really taking a long time for each of these iterations. This is going to end up taking like 30 hours. Like that’s okay. You don’t get tired. Like just keep going. And it was.

Stew (20:28)

Mm-hmm.

Jim Johnson (20:30)

I mean that sounds like

a stupid question you might get from a super new junior employee. Do you really want me to do this? This is gonna take a-

Shanti Greene (20:35)

Yeah, it was

hilarious. It was like, well, there’s gotta be a better way to like, don’t they have an API I could use? was like, no, if they did, I would have said to do it that way. Clearly they don’t.

Jim Johnson (20:43)

Stew,

I think the way you just described that though, many of the questions and frameworks and thinking you put in place for…

an employee, if they’re coming in, it really doesn’t matter if they’re coming in to help you with marketing or they’re coming in to monitor, you know, to manage some narrow portion of supply chain or it doesn’t really matter. All those things that have sort of naturally grown up through decades from a policy handbook to sort of the unwritten rules of the road that are now going to need to be written in terms of how you operate in the workplace to evaluating the quality of their work, all of that.

will need to be replicated in just the right way for your sort of independent but super

Stew (21:27)

⁓ I absolutely agree.

Shanti Greene (21:28)

yeah.

Stew (21:29)

Yeah, I think ⁓ and I think the interesting thing for humans is we ⁓ we’re all going to be supervisors. Right. ⁓ And so I think these

management skills are in terms of, you know, how to bring clarity to what it is you’re trying to achieve, how to put structure around how to break up the work and get it done, how to clearly communicate, you know, to just an intelligence around what you’re trying to do. Those are all going to be even more important.

I think in this world and more people, right? Maybe those types of supervisory skills just based on the math was something you’d expect from 15 to 20 % of your workforce. I think in this new world, you’re really gonna need to expect those skills from nearly 100 % of your workforce.

Pete Reilly (22:31)

And,

and so that.

Jim Johnson (22:33)

At some level somebody has to own.

In this world you’re describing, someone, human somewhere has to be accountable to… Accountability is huge here because there’s not an outcome where something, whether it’s a customer service agent or whether it’s a sales agent or any number of roles, where something bad happens and it can’t ultimately be owned or pointed to a human as accountable for the outcome.

Pete Reilly (22:42)

Accountability is a big question, right?

Shanti Greene (22:44)

Mm-hmm.

Stew (23:04)

Yeah, and I think that’s why I think what you are gonna start to see is ⁓ Each of us Kind of having agents we’re accountable for they’re assigned to us in the way that you decide to supervise you get a budget These things by the way aren’t super cheap

Like I,

Shanti Greene (23:22)

Yeah.

Stew (23:23)

I’m probably, you know, I don’t want any of y’all to know how much I’m spending a day on agent calls. It’s probably, you know, much as a person. ⁓ yeah, it’s, you know, it’s pretty bad. You know, you can be a couple hundred dollars a day. You could easily run up with some of these coding things. ⁓ so I think you’ve got a, ⁓ you know, you’re going to have a budget. You’re going to be,

Shanti Greene (23:29)

you

Jim Johnson (23:31)

I’ve actually seen the

Shanti Greene (23:32)

I’ve seen your Claude API bill.

Stew (23:51)

You’re going to be, these agents are going to be assigned to you when you leave, right? If you leave, well, that agent was doing important work. It probably had its own email address and had its own Slack. was tied to things that are set up. How does that transition work? There’s going to be a lot of weird stuff we have to work through. I think around the way these ⁓ non-human employees effectively work ⁓ in our enterprises.

But I agree, Jim, that I think it’s a lot of the same constructs we’ve grown familiar with and COVID taught us even more so how if you think about it as just ⁓ remote work. I think everyone getting that experience with remote work and work that’s mainly driven on a computer I think has trained us around a lot of these things.

Jim Johnson (24:18)

You.

Some of the things though that are constraints or constructs or guidelines for work.

⁓ go unsaid and we don’t even have to think about. just, you know, whether we sort of look at the schools someone came through or the degree programs or the interview process or just the culture, you sort of know it’s there and you’re counting on it. Once in a while, maybe that’s an issue and you have to take corrective action with an employee or maybe they don’t work out and you move them on. There’s a lot of things that are sort of unsaid that are going to now need to be said.

in terms of framing up how the agent will work because we’re sort of creating that employee out of nothing. The one other thing about the, so I totally agree with you on the sort of employee, the virtual employee or the digital employee idea. The difference though in this situation is so many people who maybe aren’t close to the agent concept will think about this in a one-to-one sort of relationship. And the mapping I don’t think is at all gonna be a one-to-one type of model today.

Pete Reilly (25:46)

Yeah,

Stew (25:47)

Not at all.

Jim Johnson (25:48)

And what exactly is that when we start narrow and sort of the boundaries that are crossed? It’s a very different model. ⁓ I totally agree with

Pete Reilly (25:49)

100 to 1.

Stew (25:58)

you

Jim Johnson (25:58)

the framework, but man, the one-to-one isn’t going to be where it’s at. It’s going to be something.

Pete Reilly (26:02)

Yeah, hold on.

Stew (26:02)

Yeah, I think the biggest,

I mean, that goes back to the biggest thing from this weekend that I think we all have to take away is things are going to get really weird, really fast, you know? And, ⁓ you know, and, and it’s a, we all have to, I think, kind of figure out how, how things are going to work in this, in this space, but, but it’s definitely some, ⁓ a lot of weirdness coming and it seems to be coming even

Jim Johnson (26:11)

.

Pete Reilly (26:12)

They certainly did this weekend. ⁓

Stew (26:32)

You ⁓ know, it’s always a little faster than ⁓ our brains are equipped to take.

Pete Reilly (26:35)

Ask for everything.

Yeah. No

doubt. Cool. All right. Well, that next topic we wanted to sort of shift gears. We want to talk a little bit about Claude Cowork, right? So if, I guess I would say if OpenClaw is sort of on the super automated technical sort of side of the equation, this is maybe on the other side of the spectrum, I would say. This is more targeted towards business folks. You download this Claude

⁓ It’s actually part of the whole Claude app on your local desktop. And it’s a little bit different. And also I’ll just describe a little bit of how I played with it and we can sort of pick up from there. So it was funny because the first thing I did is I downloaded this. And of course, ⁓ part of the magic, I think what they’re trying to do is take the magic associated with Claude Code with more technical folks and bring that to more business users. And so a big part of that is access to your local files and so on.

But of course, the first thing I do is all my data is up in the clouds. I’m like trying to download transcripts and do all these things and get all the connections working, which didn’t work really well at all. And then I said, oh, wait a minute, I need to reorient myself around the use of this here. And we were having an internal meeting about, and Stew, you were there, we were talking about company plans for the year. And I had this spreadsheet.

three companies in this one spreadsheet, probably, I don’t know, 15, 20 tabs or something. So I just dropped that into, I gave Claude Code, or sorry, Claude Cowork access to that. And I said, hey, I’m in this meeting, I need a little summary of what this plan says. And it did a really great job of just sort of producing just a little narrative of just the high level version of the plan. And for me, what was really valuable is that

Stew (28:16)

you

Pete Reilly (28:35)

as people had follow up questions, so Alon our CEO, he’ll often ask questions about some detailed part of the plan. Like one of the questions was, who’s on this team? And what he really dislikes when I go and open up and go find the, try to find the tab and open the tab and like do it. So it was great because I can just sit there and go, hey, what, what, know, who’s on the team? And it, and it did a great job of answering that question very specifically and just kept the meeting on track and. ⁓

Anyway, my experience with it was really good. And so I actually started trying to orient more and more around local files around even planning my actions and to do’s and so on and working on more integrating into my flow there. I found it initially it was a little awkward in terms of how I was thinking about it, but over time as I started to use it, I saw some real value from it. Anybody else? Go ahead, Shanti.

Shanti Greene (29:24)

Yeah. Yeah.

I played with a little bit. I actually think that Claude cowork is on its way to being a Claude bot and open Claude. It’s trying to do many of those things, but they’re putting more guard rails around it and not just, and not just saying like here, we did it all right now. They’re like, we’ll do it a little bit at a time. They’re going to drip it out, have a few extra releases, but yeah, I had an interesting task. Like I was creating a bunch of Claude.

Pete Reilly (29:35)

Yeah, lot more guardrails around it.

Shanti Greene (29:49)

I had just made a bunch of image files, but they were all named with just UUIDs on them. had no idea what was in them. I didn’t want to go open, you know, a hundred files and be like, Oh, how do I organize these? I’m like, I wonder if cowork can do this. I was like, let me set this to a task. Like, Hey, I just told them to rename the files directory. And it did a fantastic job. Opened, looked at them and then organize them into four groups. I realized that I had like two different sets of logos that I had been creating and that I didn’t even remember I had done because I couldn’t see them all.

And then with like this four folder structure, name them based on like the style, like, okay, these are like your digital art style. These are like a more realistic style. Absolutely fantastic. I was really happy with that type of use case, which would have taken me probably a couple of hours to do manually if I had done it. And that’s why I hadn’t done it yet. So that was really nice. And I think there’s a bunch of those types of little tasks that are just repetitive, but you need to use a little bit of judgment.

Pete Reilly (30:37)

Bye.

Shanti Greene (30:47)

on each one of the tasks and go, what is this? And Claude Cowork is going to be very good at that.

Pete Reilly (30:50)

Mm-hmm.

Stew (30:52)

I think too for enterprises, there’s a concept with this co-work of plugins basically and plugins are mostly like markdown files, just text files that sit in a directory, maybe a little bit of code, but they define ⁓ how to do certain tasks. I think it’s…

This is going to be a focus for organizations. I think it’s figuring out what are the plugins for your business. ⁓ You know, so how do you take this right now where it co-works really helpful for this sort of use cases you guys have mentioned, which is, I have this file or I have, you know, some things I want to do on my PC. Cause that’s what it’s able to know. But you know, if it had a plugin that could connect to your CRM system and then you could ask it any.

Pete Reilly (31:20)

Hmm.

Stew (31:42)

anything about your CRM data, if it had, you know, and these kind of already exist. You can get early taste of what that is in some, with some systems, you know, be in a product development meeting and you’re asking it about your, know, JIRA board or whatever, you know. ⁓ So I think that’s going to be an interesting thing for enterprises is, is, ⁓ is just kind of connecting these agents.

to our systems and giving them more of that context around the specific stuff we were talking about, which goes exactly into what we were talking about in the previous discussion, right? Like I say, this is just a little more on Rails and probably more suitable for an actual enterprise environment. But it’s a good early glimpse.

Pete Reilly (32:30)

And to me, it’s

oriented highly around local files, which was a little different for me as I was talking about earlier. like, all my stuff’s up in the cloud. when you work with code, it’s all on your hard drive, right? So it’s sort of orient. I wonder if there’s going to be sort of this, I don’t know, shift of.

Jim Johnson (32:38)

Right.

Stew (32:46)

Everything’s local.

Pete Reilly (32:53)

getting your taking away your files off the cloud and getting them locally or it’ll be really interesting to sort of, what do you guys think that is maybe headed?

Shanti Greene (32:59)

I mean, I use the Google Drive plugin for it. So there’s files I don’t have locally, as long as I kind of point it to the right cloud structure, it’ll go find things there. And I think there’ll be more of that, which are as long as you give it the right connections to find the data, it’ll go find that data for you and put things together. And we get it back to that same problem of how willing are we to say, here are all of the connections.

Pete Reilly (33:02)

Did.

Yeah.

With the mona.

Jim Johnson (33:27)

Ruh, rah.

Shanti Greene (33:27)

that you could

possibly have to all of the data sources, try not to do anything wrong.

Jim Johnson (33:32)

There’s the worry over that individually. So I’m going to give it access to my email, and I’m going to give it access to my files. In some ways, I’m still the sandbox, if you will. I’m ⁓ sort of disaster of one, potentially. But the moment you say, I’m going to

Shanti Greene (33:44)

Mm.

Jim Johnson (33:52)

plug this into my CRM and it’s sort of cutting across all users, then the sort of enterprise issues and enterprise concerns and enterprise guardrails become a bigger thing, even if it’s real.

Pete Reilly (34:04)

I can go

to the web and use Claude there and do the same connection to Google Drive there. Okay, so then what’s the point of Claude Cowork, which is sitting locally?

Shanti Greene (34:16)

that can connect both of those things and you can take action on things that aren’t network connected. if, yeah.

Stew (34:22)

I mean, connect your email

and say like, Hey, can you go through and give me a summary of all my emails from Jim over the last week? ⁓ build it into an action item list. ⁓ you know, things along those lines with the customer. Yeah. I mean, I think those types of tasks connecting to the system are already very valuable. ⁓ with, with, ⁓ like the co that’s probably one of the things I will use it for is just say like, just, you know, remind me like.

Jim Johnson (34:32)

Go read through the last 20 days and show me what’s important and what I may have missed.

Stew (34:51)

What were the action items coming out of my last three emails or, Hey, I got six different threads from around this one customer project. Can you just put them all together and tell me what the heck I’m missing out on? You know, it’s actually good at that.

Pete Reilly (35:03)

Mm-hmm.

Shanti Greene (35:04)

Yeah,

related to that one, are you guys using the new Google Labs feature Gmail CC, where it’ll summarize and at the beginning of the day give you a here’s the things you should be doing today by, it’s great. Yeah, sign up for that one. I mean, it’s another email that you’ll get every day, but it’ll be a summary of here’s what you should be doing today. And it does a pretty good job, except for if you get enough advertisements that hit your inbox, it’ll be like, hey, you need to make a decision about this thing. Like, yeah, the decision is I never wanted it.

Pete Reilly (35:11)

You

Stew (35:16)

I haven’t used that yet, that’s cool.

Nah.

Jim Johnson (35:34)

I’ll

be going back and summarizing the last 10 summaries, a pyramid of summaries.

Shanti Greene (35:38)

Yes.

Pete Reilly (35:38)

But

again, you can sort of do a lot of that in the cloud version. It sounds like, in the scenario you were talking about, it’s actually creating a file as output to that on my local machine. Is that what you’re saying?

Stew (35:54)

No, I mean, you can, mean, with, with cowork, you can, ⁓ you can still connect to these other systems. ⁓

Pete Reilly (36:01)

That’s what I’m saying.

That’s what I’m trying to make sure we’re understanding is, okay, I can connect to these, you know, Gmail, can connect to a Google Calendar, can connect to a Google Drive, I can do that locally or in the cloud. But it felt to me like, okay, there’s a point, there’s a reason they want to do this locally. And Shanti, you talked about being able to interact with those files. And so the scenario you were describing, Stew, I thought, yeah, the benefit of being connected locally is I could then create

Because today, if you go to the web app and you ask it to create a file, it creates a file and you have to download it. Here, it could actually just create a pattern. ⁓

Jim Johnson (36:34)

I I saw

Stew (36:35)

That’s true, right? Yeah.

Jim Johnson (36:36)

co-work as a narrow, sort of as a narrowing and extrapolation away from Claude code for those people who are sort of intimidated by the Claude code interface. And it, in essence came with sort of, here’s a series of things you can do out of the gate and the community can sort of add to it.

or you can personally add to it. ⁓ But ultimately it’s a subset of what you might could have done already with Claude Code if you wanted to. Ultimately, they’re sort of the same thing. It’s just that one was an easier, smaller sort of set of use cases initially, but you could expand on it if you, you could sort of add to cowork yourself or you could borrow from other people who are adding to cowork. Maybe that’s a misrepresentation.

Stew (37:28)

I think that’s right. It’s kind of a… ⁓

a new UI for Claude Code trying to bring that capability to normies.

Shanti Greene (37:39)

Yeah, of course,

Pete Reilly (37:39)

to

the local.

Shanti Greene (37:40)

bringing in another like nomenclature for now we’ve got skills agents, subagents plugins, which are really just a combination of skills and subagents. At some point, I think we just need to like clear up the naming. We sort of like what, happened to MCP? Well, people are like, well, we don’t need MCP. Instead, we’re going to use agents and skills. And it seems like that we’ve definitely kind of gone away from MCP towards the skills.

Jim Johnson (37:46)

plugins.

Pete Reilly (37:55)

Bye.

Shanti Greene (38:06)

And now it seems like subagents are actually more powerful than the skills, according to like Vercell’s article and the testing that they did. Like, no, we should stick with the subagents as long as they’re well defined. So figuring that out and it’s moving quickly, which is nice.

Stew (38:19)

Well, I think the other thing, you were mentioning about this happening on your PC, which is how CoWork mostly works today. know, Claude Code already can run in the cloud, and it launches a sandbox environment. But it uses GitHub basically as the glue to this whole system.

Pete Reilly (38:36)

Right, as the common preference.

Stew (38:40)

I think that’s probably going to be, if you combine like what we talked about earlier with the Claude bot and then with this, think the next evolution of this, which will clearly happen is this will be launching a sandbox environment. And then they’ll need to be, you know, then maybe that sandbox is using your Google drive or, you know, something like that as a, as a common place for the artifacts it creates.

Pete Reilly (39:02)

Yeah.

Stew (39:09)

to come back to things along those lines. And you’ll have access to it on your phone and all of those other sorts of things. And maybe it’ll be able to access things on your PC. So it’s an interesting vehicle, early days for sure. But yeah.

Pete Reilly (39:24)

Yeah. Now I expect we’ll see more of those from the other foundation model

guys. Okay. So let’s, let’s shift on to the third and final topic. We last, in the last episode, we used a CRM as a way to sort of just demonstrate a couple of things. How quickly it is now to build software. And also this concept of, well, if you don’t, you don’t necessarily want to just create another

you know, create, read, update, delete, crud, you know, ⁓ CRM. And you need to make sure that your sort of AI is deeply embedded in facilitating the human’s work. And that sort of then brought us into this other debate, you Vivian and I were having about, you know, should you or should you not, or when should you or when should you not, you know, use all these capabilities to create your own CRM. I mean, I assume

You guys have all seen the stocks of Microsoft, Salesforce, and others basically tank despite the fact that they were delivering ⁓ great financial results. And part of it is this mindset of, wait a minute, everybody’s going to sort of build their own ⁓ CRM now and their own software. So let’s tee that up. I know who wants to pick that one up and give your initial perspective.

Shanti Greene (40:48)

Yeah, I’ve got some thoughts because of how easy it is to start building something and how difficult it is to end up with a polished product. But if you’re looking at building something for your internal use and it doesn’t have to be as polished, you might be able to get away with 90 % of the features and that starts to push you more towards, maybe I should build this.

Pete Reilly (40:56)

Hmm.

Shanti Greene (41:08)

The other big thing is I think we all basically as individuals or companies have subscription fatigue. I don’t want to pay another monthly per seat license for whatever the thing is where I can pay tokens upfront and have a thing that I can use. However, I see fit not have to pay extra when I want that one extra feature. We’ll build it in for what we need and it’s built to purpose. Now it’s not extensible.

Stew (41:17)

you

Shanti Greene (41:37)

you’ll have to basically rebuild or have people to build those features and you have to support it. So there’s lots of ongoing costs. They’re just different ongoing costs. And depending on how you want to manage that, it could be a good option. But if you’re going to now turn that around and make that, we want a customer facing or we want people outside of the development community to use it, maybe I wouldn’t lean so much towards rolling my own at that point.

Stew (42:04)

I think it’s, know, this using third party software versus building in houses always kind of come down to opportunity cost and differentiation. Right. And this, I don’t think this changes that it’s just changes, you know, maybe what those costs are or what the opportunity is to differentiate. So I think, you know,

Shanti Greene (42:25)

Yeah.

Stew (42:29)

First thing to do is like, well, you know, could I build my own payroll system? ⁓ Yeah, but the energy I’m gonna put to that, is ⁓ energy maybe I could have applied to something that’s much more differentiating for my business, right? So I think where you wanna, you know, where I think businesses will find the most value in ⁓ using these tools to build apps for themselves versus applying third party apps.

Is when it’s really hitting at the core, what differentiates their business from other businesses, right? Cause that’s where you want to apply your energy is what really separates me from the crowd. If it’s just to get, you know, whatever and, ⁓ software that does something exactly the way it’s done for a zillion other companies, you might be able to do it faster. You might be able to save a few bucks on a subscription. You might want to.

Shanti Greene (43:24)

you

Stew (43:28)

You might be to do some of other stuff, but the opportunity cost of doing that for little to no differentiation, think it’s going to still offset the time it requires for you to do that. What I think will happen is probably deflation in the SAAS. In other words, it’ll be easier for competitors to come and suck the margin out of the other third party tools that you’re using there.

less, you know, so I think that’s what’s going to be interesting. And maybe that’s what’s driving some of this SaaS impact on the stock is, is less that everyone’s going to build their own CRM, then it’s going to be easy for new entrants to come in and build and pick off little niches and, and shave people away.

Pete Reilly (44:10)

They’re going to pick off little niches where, you know.

Shanti Greene (44:12)

Mm-hmm.

Jim Johnson (44:13)

Well, yeah, I would sort

of think about a couple of thoughts.

⁓ one sort of in a satellite model, I think we, know, even in a, even in a Salesforce world or an SAP world, everybody doesn’t get what everyone wants today. regardless of the fact that, you know, 150,000 man years have been spent building Salesforce by Salesforce, you still don’t have everything you want. So there are sort of this idea of, of sort of feature at systems or capabilities that people are going to need to build, be able to build around the.

edges now. The other thing that I would argue that sort of is a maybe a supporter of that kind of outcome

⁓ especially at the enterprise level, I believe that it and the business, if I could sort of just simplify the world to that have become increasingly divorced over time, whether this is because of sort of large it service models with the big players, ⁓ with, onshore offshore, just a whole bunch of reasons that, that they sort of become divorced from one another. And you have pent up demand.

Shanti Greene (45:11)

Mm-hmm.

Jim Johnson (45:27)

at the business, you know, they’ve tried to solve it over the years with the server in a closet or shadow IT or whatever. This facilitates it more than anything else that we’ve seen in a long time. I don’t think necessarily that you’re going to see someone building Salesforce or SAP as a cost management framework. And certainly you don’t want to take on your own payroll, but sort of developing satellite features and capabilities that sort of add around

the edges for what you want, for what you need for your business. I think that’s here and I think we’re going to see more of it. And I think some of it’s manifest with it or it was a manifestation over frustration in that, in that gap between what the business really needs to deliver value and sort of where IT has evolved to sort of standoffish and far away from the business. ⁓

Pete Reilly (46:19)

Mm-hmm.

Stew (46:19)

I think so much of that IT standoffishness is really comes back to, ⁓ constraint.

You know, basically like, right? Like, I’ve only got so many IT people. I’ve only got so much stuff. So you put all of this infrastructure around reality is a lot of that stuff’s easy, right? A lot of stuff that was hard is now easy. A lot of stuff that that was hard is now harder. You know, I mean, it’s just different. so I think there’s a whole, the organization where I’ve got to go to IT and get approval, you know,

Jim Johnson (46:33)

Yes.

Pete Reilly (46:39)

budget.

Shanti Greene (46:51)

Yeah.

Stew (47:02)

because, ⁓ well, there’s only so many IT people and we can only take so much work and we only have to do whatever. Cause it would take six man months to get this thing done. Well, if it can be done in six days, right. Instead of six man months, you’re going to have a much different, ⁓ philosophy around how this works. Yeah.

Jim Johnson (47:17)

It’s days on its technical understanding though. It’s

previously there was like, man, the wizardry of what’s happening behind the screen at IT, we can’t touch that. So yes, it’s days and budget, but I also think there will be less fear over taking it on right now.

Stew (47:38)

Great, yeah, I agree, yeah.

Shanti Greene (47:38)

Yeah.

Pete Reilly (47:40)

Yeah, the

build versus buy, and we said this in the last episode, the build versus buy line has absolutely moved. And I was just sitting here thinking, oh, it’s not just build. It’s it’s build and maintain, right? Because the minute you put something out in the world and you start getting users, well, now you have to evolve it. You’re going to have bugs. So that has to be in the equation. But it’s absolutely moved in terms of what just happened over the last, and I’ll call it,

call it two months in terms of code, but being able to write software and so on. I do think, I don’t know about you guys, I’m building all sorts of personal software to just help my personal productivity, right? So I think you’re gonna see that ⁓ continue to evolve and that’s not actually gonna be replacing a CRM. There are situations though, Jim, you’ve seen this in the consulting career where, I mean, I know we’ve come in and folks have said, we want…

You know, we have this custom thing. We don’t want to support it anymore. We’re going to go with Salesforce. But then the very first thing that ends up happening is there’s a huge tab to customize Salesforce and that’s the way they run their business. they would constantly push back. The customers would push back and we’d say, listen, if you’re going to go package software, you’re going to have to change your processes to match the package software. Unless you want to go fully custom, but that’s what you’ve got and you’re not happy.

Shanti Greene (48:44)

Mm-hmm.

Pete Reilly (49:01)

So there’s still that interesting dynamic and there may be situations where if it’s going to cost you a million dollars to sort of customize something, well then maybe the line has moved and you should build it. The other thing I would say is if you’re just going to build sort of the traditional CRUD tool that probably isn’t going to make a lot of sense over coming years, really, if you are going to go down that path, you have to design it with an AI first mindset.

to really have something that’s going to provide value and survive over the next ⁓ couple of years. And ⁓ one question I was going to ask you guys is that, do you see a world where something like a Salesforce still acts as the system of record, but people start building their own apps sort of around that to meet their custom workflow, but not necessarily replace the core.

Stew (49:57)

I mean, think this will be the differentiator for those type of companies is how easy are you making it for your software to work with these agents? And so I think the existing SAAS provider should have an advantage because they have

Pete Reilly (50:05)

Mm-hmm.

Yeah.

if they embrace it.

Stew (50:21)

They have all of these systems,

they have APIs, they have all these sort of things that are the system of record. have the installed customer basis, things along those lines. But I think if they wind up kind of trying to protect all of the services revenue, right? That they’ve had before or try to put up certain gates. I think that’s going to be the opening for new competitors to come in and say like, no, I’m the AI. I’ve saw something in us today. Like I’m the AI first CRM.

Pete Reilly (50:49)

Quite this.

Stew (50:49)

⁓ They

got a big new company, forget what they were called, launched today, that had a bunch of funding behind them. And their whole thing was basically like, we’re going to be the ones that are easy to work with your agents, easy to do the type of things you’re talking about. So I think that’s where the competition is going to be and where we’ll see how well the legacy SaaS providers step up to the game in terms of making their software ⁓

easy to work in an AI native world versus new entrants coming in that are just designing for that ⁓ use case from the beginning.

Pete Reilly (51:23)

and

Shanti Greene (51:29)

One of the things I look for when I’m evaluating products, like do I want to use this now is do they have an API that I can do all of the operations with? Do they have webhooks so I can get in near real time that something happened in that system? Part of it is, you know, want to have access, want to make it easier for agents, but also if I need to migrate, will I have access to all of the data that I can recreate everything I had?

Jim Johnson (51:53)

Exactly.

Shanti Greene (51:53)

And if

you don’t have those pieces, why would you adopt new software that’s going to try to like lock you into something? Whereas I need access to all of my data because that’s what’s key to whatever it is I’m doing is the data, not how I get it or what I’m doing with it.

Jim Johnson (52:07)

And how that data is constructed underneath or laid out underneath what the data model is and what the metadata, the semantic layer, and maybe even when we start talking about ontologies and how that really lines up to your business, that is sort of…

Shanti Greene (52:10)

Mm-hmm.

Yep.

Jim Johnson (52:23)

enabling access to it for agents. And we don’t necessarily know exactly what the surface area of the agents are going to be. It’s not going to maybe be in the traditional roles. Maybe the UI doesn’t matter that much on the SAS applications. Maybe it is much more about the underlying data and how it surfaces it to the agents. And we won’t really care if it does the UI mirror a human workflow.

Pete Reilly (52:48)

I guess the answer to the question, should you build your own CRM is maybe, it depends. ⁓

Shanti Greene (52:52)

Maybe, yeah. ⁓

Jim Johnson (52:54)

Well, but

just one last comment on that. There’s this interest, there’s a spectrum of, you know, the smallest company, what is their CRM? It’s probably a spreadsheet.

And then somewhere along the way you decide to jump up to one of the next level CRMs, there’s probably a space in the middle where maybe you do. It doesn’t mean that you’re sort of necessarily replacing the jumbo CRM, but moving from just a basic spreadsheet to a full-bodied CRM, I still think there’s niches in there that this may need to.

Pete Reilly (53:30)

and I fully expect

a lot of startups to be the, you know, I’m the AI CRM for plumbers. You know what I mean? Like it’s going to be, there are going to be these niche sort of offerings that do make sense as sort of a smaller, a smaller business and not really go up and compete against the Salesforce because they’re not really serving that audience. I can see that absolutely happening. Any parting, parting words of wisdom?

Shanti Greene (53:54)

No, people should try the new tools. They’re there, try them out and try them in a way that is safe for you. Sandbox it, give it limited access, but try them out, because you need to know what these things are capable of.

Pete Reilly (54:00)

Yeah.

Jim Johnson (54:01)

If

Pete Reilly (54:04)

You need know what’s going

on. Yeah, that’s a good point on that. We’ll close it out. Guys, thanks for joining us. Have an awesome week.

Shanti Greene (54:12)

All right, see you guys, bye.

Jim Johnson (54:13)

Thanks gang. Happy