← Using AI at Work: AI in the Workplace & Generative AI for Business Leaders
Using AI at Work: AI in the Workplace & Generative AI for Business Leaders
113: Using AI Agents for Business: From Chatbots to Autonomous Workflows with Bryan McAnulty
00:52:25
Loading summary
Transcript59 lines
- [00:00]
A
The seven hours that the agent spends, it's not seven hours of human work. A lot of times the equivalent of what that actually did would be maybe months of what it would take a human to do.
- [00:09]
B
So how does somebody identify new task versus new chat?
- [00:13]
A
There are many things that we have to do day to day that involve like research or like analyzing and transforming information in some way. And so that's all the stuff that you can offload to AI.
- [00:24]
B
Can you define the difference between what you would call an automation and something more along the new task conversation?
- [00:30]
A
The automation is basically just the recurrence.
- [00:33]
B
What does bigger customer outcome look like in practice?
- [00:36]
A
If an agent does all the work, then the processes or the way we do it is the thing that matters. So inside our platform latch loop, our approach is that like the general knowledge worker agents, all the memories and everything that's in a GitHub repository, you can take it anywhere, use it with whatever tool you want. The way we're going to be competitive is I hope we're going to provide a better workflow and an interface that you prefer to use our tool. But if not, like we're not going to hold that data hostage. Using agents wrong would be worse than just not using them at all.
- [01:02]
B
Brian McAnulty is currently the founder and product director of Heights Platform and Latch Loop where they're focused on helping creators and knowledge businesses build and grow online. Welcome to Using AI at Work. I'm your host Chris Staigle. Each week we'll be learning how today's business owners, entrepreneurs and ambitious professionals are getting more done with smart use of tomorrow's tech. Let's get started.
- [01:25]
C
Right now, every business leader is asking the same question. What are we going to do about AI? If this is you, chiefaiofficer.com has the answer. We give you a simple path forward where we provide executive and team training so your people know exactly how to safely use generative AI in their day to day. We also manage the deployment and implementation to make sure tools actually get adopted and deliver results. And we'll also guide company wide transformation so AI becomes part of your operating system, not just another shiny object. The companies that act now will increase productivity, cut costs and grow faster than their competitors. Those that wait will get left behind. So if you want to make AI work in your business, visit chiefai officer.com and see how we're helping companies of all sizes finally get results from AI. Hey everybody.
- [02:12]
B
Welcome to Using AI Work. My name is Chris Dagel and I'm the host and today Our guest is Brian McAnulti. Now, Brian and I are both Austin area AI entrepreneurs, and over the course of our discussions, we identified that, or I specifically identified that the way that Brian is thinking about and deploying AI agents is something that I think will be kind of the trend as more and more people get on it. And I have invited him on here today to kind of share that perspective of kind of what that transition looks like, what he's been up to, how he's been leveraging them for internally and for clients, and so that you can walk away with, I think, a better foundation from which to make your decisions on what are we going to do about agents in the organization. So Brad's currently the founder and product director of Heights Platform and Latch Loop, where they're focused on helping creators and knowledge businesses build and grow online. And he spent more than a year developing the AI agent platform that we're going to talk about today. So, Brian, before we get into the deep discussion, anything that you think that the audience needs to know about what you've been up to, your perspectives on AI and agents?
- [03:22]
A
Yeah, thanks so much for having me on, Chris. Thanks for being here. Yeah, I appreciate it. I want to kind of share with everybody where we see this going because this is something that we've been working on from the start. We've had these ideas of things that we wanted to do in business, but they just weren't possible. And then GPT4 came out and it changed everything. It finally made all these ideas, like, wait, now we can use AI to accomplish these things. And so I was really like digging into all these things right away to see, like, how can I help my small team move faster with AI? And with Latch Loop, it really started as an internal tool, just like me being fed up with what was available and at the same time seeing what I knew was possible. And so that's why we started building it. Like, how can we build our own products and software faster? But I think we stumbled into some things that I believe is a better way to work with agents. And so whether people use my platform or not, I want to help prepare them for the way that we see the kind of collaboration between humans and agents in the future.
- [04:35]
B
And we've got a number of kind of, I guess, topics that I want to address on this, but just so that the audience has some context. What is your background? Is it technical? Is it operational? Is it entrepreneurial? Is it more traditional?
- [04:48]
A
Yeah, always been an entrepreneur, never went to college. Right out of high school, I was Running a freelance business, turned into a web design studio. Eventually got fed up with working with clients, turned that into how can we build some of our own software, started building SaaS products. The most well known of those would be Heights Platform where We've helped over 10,000 creators build these online businesses. And so my specific background originally was not as technical, it was more so like a designer. So I still see myself more as like the product designer than a maybe like traditional software engineer. But over time I've had to learn all the code and I've been building all these things. So we're a bootstrap company and so we try to be like very lean and efficient in how we handle things. But yeah, in terms of how that's led me to build what I'm building today, when I saw that people were working with these coding agents as a chat window in an IDE or a cli, to me it was just absolutely ridiculous. It's like you're telling me that we have this technology of the future and the way we're going to interact with it is the interfaces we had from the 80s. Yeah. And to me it's just like that, that there has to be a better interface for us to engage with these things. So that, that was the mission and actually like trying to bridge the, the gap of like how can we make this approachable for non technical people to interact with these agents and how can they make it so my team can comfortably use these agents and be able to contribute to our software and things like that without feeling that like, oh well, have to ask a developer first.
- [06:34]
B
Okay. So I know that, that generally your thesis is that most of us are using AI and we're pushing the new chat button. Right? We're, we're chatting with AI as compared to your perspective is that we should reevaluate how we're approaching this and look at it as new task. Right? Yeah. So that is a fundamental change from what, um, you know, generally what you're seeing out there as far as like, oh, these are the skills that you need to have to be an effective knowledge worker or business executive using generative AI. It's prompt skills, it's maybe it's something as, you know, basic as connectors and that sort of thing. Right. But the way that you're seeing it is because the chat, I've said this before, maybe even on the podcast, but it's like a drive up window. Hey, I need an answer. Got my answer, I leave. And what you're talking about is a completely different as you just referenced, like it's, you know, the most powerful technology we've ever had available and we're drive through window. Give me an answer, I'm out of here. As compared to what you're suggesting is a different way to approach it from the task level. So can you kind of explain that, that thesis a little bit more for the listeners?
- [07:46]
A
Yeah, so, so when we built this, this agent platform latch loop, the approach had always been that you're not chatting with the AI, you're creating a new task. And the interface looks much more similar maybe to like a project management system where you have this like task document as opposed to just a long chat thread. I think this solves a couple problems in like the UX and what a business actually needs to see. But before I get into that, like the fundamental difference of why we're creating a task and why I'm really excited about this because like last Thursday I think ChatGPT, they released GPT 5.6 and the ChatGPT app has been completely overhauled on desktop to now be this ChatGPT work plus ChatGPT Codex. So Codex, the AI coding agent is merged into the ChatGPT app now. And now in the ChatGPT desktop app, the default is new task. Now this confused a lot of people, but for me I was like, well, I've been doing this for over a year and so I'm excited for this. For me I feel like it's educating the market and what we've been trying to build. Yes, but the idea is that these agents can work on these much longer running tasks. And so the way that we need to think about interacting with them is much different than the idea of what you said, like the drive thru window and like, hey, I'm going to send one prompt and then we can get one answer. And for not only for like founders and business owners, but especially for like employees in teams, the way that they work is going to become very, very different than what they're used to because it's less about, hey, I just need one answer, as if you would Google it or something like that. And more about like, how can I now learn to delegate something to these agents? And there's things that have to be learned not only in the processes of how are we going to actually work with these things and what goes to the agent versus what do I do? But yeah, also even the understanding of, well, what are the things that I should always be doing versus what are the things that the agent should always be doing? And to give like one. Yeah, One more clear example, I guess, like, we have these agents now working on, like, tasks that they'll just go for, like. Like seven hours or so working on that one thing. And then sometimes you may want to, like, steer it a little bit, or maybe you have to approve something, but oftentimes, like, it can just go off and work on that thing. And the seven hours that the agent spends, it's not seven hours of human work. A lot of times the equivalent of what that actually did would be maybe months of what it would take a human to do. And so I think most people are thinking too small. And a real problem that exists in these tools right now is we have the same chat input that we had before, but the models and the agent harnesses and everything is so much more capable. And so you kind of have to think bigger and try to experiment to push the limits of, well, what can we actually do here? And using these things all day, I'm still constantly surprised at what they can actually do.
- [10:53]
B
Yeah. So a couple things to unpack for the listener. So I don't know if you caught that, but Brian was referencing that the ChatGPT desktop app has been combined with Codex, which was Chat GPT more, I want to say technical, but more robust. It allowed you to do more technical things than just the chat environment. And I don't know if you caught that, but he's been talking about this idea of new chat versus new task for a year internally with his team and with his clients and that sort of thing. But now that's an option with the Chat GPT desktop. So what he's been kind of anticipating occurring like this is it's happening now commercially mainstream. You now have access to it. Secondly, this idea, one of the things, Brian, is that when we, for our services side of our business, we talk to executives and they're like, we'd love to use AI more, but we don't know where to use it. Right. Like the, the conundrum of the use case. Where do I get started? Now, that's for basic stuff. Right. That's not for agentic task work like you're talking about. So how does somebody identify, hey, this would be an opportunity for me to click on new task versus new chat.
- [12:04]
A
Yeah. So, yeah, let me also unpack a couple things about what I'm about to share is that I think a big unlock for a lot of the companies building these agents and the model providers themselves was that the coding agents everybody's been working so hard on because we realized the AI can work on code for a long time successfully here. The coding agents are actually the same as the general agents. If you went back like a year or two ago, I think everybody had the mindset that like, okay, well we all have to build these different, like AI agent softwares for these different maybe verticals and things like that. And in some ways I believe that to be true based on like the data in a very specialized vertical. But in reality, like the big unlock has been that the general knowledge worker agent that an employee could use in the company who is not technical, is internally the same technical foundation as how the coding agent works. It's just instead of adjusting the lines of code, it's adjusting lines in a Word document or a PowerPoint or a spreadsheet. And now so when it comes to what does somebody actually, when you're trying to understand how to work with these things, what does somebody actually decide to delegate to AI versus what should they do themselves? My kind of extreme take on this is I've jokingly, but not jokingly said to my team, I'm seriously considering to say that nobody's allowed to do work anymore, that the AI has to do the work. Your job is to come up with the ideas and be directing this thing and to be maybe more concrete on that. There are many things that we have to do day to day that involve like research or like analyzing and transforming information in some way that is not about the original like ideas or the taste and everything that we have as humans. And so that's all the stuff that you can offload to AI. And I think work is going to be separated into almost like three, three different sections of your workflow. Okay, one is like, what are your kind of like agentic loops? Agentic loop is like the, the process where the, the AI is able to just work on the same kind of goal continuously either. Like, I guess what I'm referring to here more specifically is like an automation where it's going to happen at some certain cadence every hour, every week, whatever. An example of that could be something like, okay, check the recent changes in our software that we released and make sure there's no security issues. And that's something like it doesn't need a human input really. You know, some kind of benefit can be had by just the agent continuously doing that. So setting up the loops and the kind of like automation of your factory of these agents is one part. Next part is setting up these big long running tasks that the agent is also kind of running in this loop, working towards a Very specific goal. And so that could be something like we had for one of our products, the agent go and research all of our different competitors. And we had it come up with 150 pages about all these different competitors, competitors we haven't even heard of. And research and well, research each one and then put out all the facts in a certain format that we defined. And that would take a human a very long time to do. But it's also very tedious work. There's no reason that a human has to be putting creative input into that. And then the third thing is the kind of back and forth where it does require that tighter feedback loop of you asking an AI, hey, can you do this thing? And then you need to see it and then say, okay, well let's adjust it a little bit like this, or changing it like that. And so building in these three types of ways and then understanding where does it need like the human ideas and taste. Because that's what the models don't get. Well, the models can do this, like this, research tasks, things like that, transforming your data, but don't ask them to come up with the ideas.
- [16:15]
B
Okay, so you mentioned automation. Can you define for the listener the difference between what you would call an automation, even though it might be an agent running that automation, and something more along the new task conversation?
- [16:32]
A
Yeah, well, I think for people using these tools for work, a lot of what you're going to do is kind of defining these tasks. And you can think of it like the same way you would work in a project management system that you have all these tasks that you need to perform at some point during the day or maybe on a recurring basis. It's just that the work done to make them a reality can now be performed by the agents instead of the humans. And the automation is basically just the recurring task. So what's the thing that you want to either check and maybe create a report about every week or what's the thing where you need to. Yeah, the. Anything that's kind of recurring in your business, that's where you can basically schedule these automations, which is basically going to wake up an agent at that time to perform this certain task.
- [17:23]
B
Good. Now you would, you would, when you say wake up the agent, that just reminded me about kind of what we talked about on the pre interview with a better understanding about these agents. And the way that you explained it was not the way that I had understood how an agent kind of operates, engages and waits for that next task. Can you, you know what I'm talking about?
- [17:46]
A
Yep, I know exactly what you're talking about.
- [17:48]
B
So do you mind explaining the audience? Yeah.
- [17:51]
A
Yes. Yeah. I'm sure a lot of you have used AI and have gotten to the situation where you think it just didn't do what I wanted, it stopped short, it didn't listen, it made something up. And I think for people who are less technical, sometimes it's very difficult to understand, well, why did that happen? Is this technology not good yet or what's actually going on? And something that I found helpful as an analogy to make a comparison here, because these agents can do some of the same work that humans can do, but the way that they work is very different. And I think that typically when we all think about AI, we're imagining this sci fi scenario that the AI is constantly thinking and running in the background somehow. And when ChatGPT came out, we realized that, okay, well, what actually happens is you send the message, then it's processing as it responds and it's done. It's not doing anything at that moment. Now we have these AI agents that seem like they're more continuously working, but the reality is that it's kind of in a way like a programming trick. And so the analogy I would give is that when the AI is responding to you, it has like come to life. It's got your, it's like you're bringing an employee to life for a moment that they have this information of the task that you told it. And the model providers have trained these models to basically try their best to accomplish whatever you said in that short moment of time. Now the problem is an AI agent is basically just having actual like programming code in the background saying, okay, after the model tries to use a tool or responds, then let's force it to run again and see what it's going to respond with next. So the analogy is the model is actually kind of coming to life and dying over and over and over. And in the process, every after every tool or message or something that it sends, it's coming back to life, but it's not coming back to life as the same model. It's technically, it's a new. You can think of it as a new entity that just got force fed. Like all the memories of the past. One was said like, hey, figure this out, what's the next step? You've got a couple minutes here. And I think when you can kind of visualize it that way, of these different little AIs coming to life and dying over and over, sounds a little bit grim, but the idea is that they're basically trying their best to work with whatever they have. And so if they don't have the right context about what they're trying to accomplish, that's where they can kind of go off the rails and say, well, I guess, I guess this is what I have. I guess this is what the person probably means by this point. So here you go. And so, yeah, the importance of having the right context for the AI and having the right tools so it can access the context to be able to do its work properly is what's going to give you the better result.
- [20:36]
B
Okay, so this is a big distinction, dear listener. You've gone from chat where you're directing things, right, and it's getting context dynamically. Here's what I need you to do. Act as a. Here's the report that I need you to review. Whatever those, those things are now in the agentic environment. You know, I, I didn't realize what Brian was, what he just shared with you. I thought the agent was just on and it was idle waiting for me to come at it. But the reality is that's not the case, that a new instance or a new, I don't know, discrete cycle of, of intelligence gets spun up when I call on that agent's instructions and that sort of thing. So if you're trying to use the agents the way that you use chat, it's obvious that you're going to be underutilizing the tool but also not getting the best results from it. So when you mention this stuff in this environment, I think this is kind of like an architecture question that I personally am interested in. But I think that will help the listeners understand that bigger picture, like the distinction between thinking about things as an agentic activity versus a chat activity. You mentioned that the agent comes back up. Where is the context that it needs so that it can pick up the conversation essentially from that last interaction?
- [21:57]
A
Yeah, so technically, what developers are doing and how the model providers have designed this is we're putting all the previous context of the conversation and everything that happened back into the agent. So technically it has everything and it is continuing. But I find that just a useful analogy to imagine that it is actually a new instance of it running. And so it's like a new soul with the same memory is put back into it, rather than being something that's constantly active. So what that means is if you send a message to a chat and it doesn't do what you want, you might find, okay, well, if I think about this, I probably didn't explain this thing, or I probably didn't give it this information to make this decision. And the analogy that I used to give people of like, working with a tool like ChatGPT is it's the equivalent of like walking up to a random person on the sidewalk and saying, hey, do this thing for me. They have no idea about any of the background context of like, if they were at least an employee in your business, at least they know, okay, I'm working for a company, the company has this general goal of doing this. Um, I probably want to keep my job, probably, as opposed to just some random person on the sidewalk who they probably want to be genuinely helpful, but they have no clue why you're saying that. And so you can realize where you need to give that extra context. But when you want to have a task that's running for a longer period where the agent's going to be working on this for potentially hours, you want to prepare it with the information or the tools that it needs. And because then it's more likely to maybe like, go off the rails or not give you the result that you want. Right. So what that means nowadays is like the connectors, the plugins, the MCP tools, making sure that your agent has access to the kind of data that it needs to perform its work. So at a simple level, that could be things like web search, it could be accessing your code base, accessing your project management system, and then that way, as long as you've made it clear, like, what is the goal that you're looking for, why are you trying to accomplish this thing and what's expected of it, it can go and kind of build the rest of that context without you having to sit there and come up with this hundred page prompt to make sure it has every little detail.
- [24:14]
B
Okay, this is great. And I don't know if this is, if you've got an answer for this, but this just happened. I was speaking to a Vistage group in Chicago. One of the things that I like to do, just like you said, the connectors, great place to start, allows for the agent to have a lot of context of what's really happening in my Slack or my email or my calendar, whatever. Like, that's, that's real actionable data points for AI to have access to, to give me direction or answer questions. The. But a couple of the folks that were in the room were like, hey, wait a minute, this seems risky. What? Like, I'm now giving AI access to these context repositories, AKA my inbox, my calendar, I Don't know that you've had that, that challenge from any clients or anything yet. But what's, what's your perspective? Yeah,
- [25:11]
A
I was surprised when in the beginning of the year we had the open claw came out and everybody's talking about that. Right. But for me that was always like a huge, huge security hazard. Of course I cannot do that in the business. And so to be clear, I'm not suggesting that anybody be reckless and say, hey, give AI access to just do anything on your computer. And I think one of the important points to understand there is the way I would think most businesses should be handling this is there should be like programmatic and deterministic guards that are fully preventing the AI from doing things that you don't want as opposed to just telling the AI in your prompt, hey, probably don't do this thing. Because again, I talked about how AI can do some of the same work as humans, but it works in a different way. If you imagine you have an employee, they're going to think about things in a different way as far as what they're going to do or not do because they want to keep their job. The AI has no job to keep and it wants to be genuinely helpful. But if you didn't tell it, hey, actually that database is the real production database. It's not just a test thing and it's just trying some things out and it has no like back and forth direction from you. It might say, you know what, they didn't make it really clear. I guess I'm going to just try to delete this, you know, and obviously that would be really bad. So the majority of these like agent platforms nowadays when you have your plugins, your MCP tools and connectors, you're able to set permissions on what kind of data is going to be available, what kind of tools are going to be available, and what kind of tools are going to require some kind of approval. And I think this is important to do and the way I do it is we have all of our different projects, so the different agents that may be a general knowledge worker agent versus a software repository, and in those each has different tools, different access, and then each has different things where it's either able to do without requiring approval or it does require approval. So looking up your calendar info, maybe you don't need approval for that, but actually changing your calendar events, you probably want to approve that before you want the agent just messing around.
- [27:22]
B
Or send, maybe just create the email, keep it as draft until let the human review and send it.
- [27:27]
A
Yeah, exactly.
- [27:28]
B
Okay. And as far as like, you know, security breach or anything like that, you, you feel comfortable as long as my. I'm using frontier models that everybody knows Claude. I mean, I guess it's already baked into Gemini or Chat GPT that, that the cyber security risk of that is minimal.
- [27:51]
A
I think it depends. You have to be aware of, of where, what your kind of exposure is and how these things are working. If they're working internally only in your internal tools, you're generally safe. If you have them like looking up information on a website, you're probably safe. But like there is this new like potential attack vector of what if somebody makes a prompt injection on a website that looks like it's supposed to be the info and then it's telling your agent, hey, actually do this thing. The modern like frontier model is if you're not using some little tiny like open source model, the modern frontier models will see right through that for the most part. So generally that's not a concern. But as soon as you have the agent kind of going out there on the web and able to interface with things like your email or other things like this, that's where it could potentially become a problem. So like imagine now we have not just phishing emails, but imagine like someone sending a phishing email pretending to be, you know, to your agent and you allow your agent to read your email. So like these are all things where you need to be careful and that's the reason why you wouldn't want the agent to send the email. Right. Let's say the. You have your agent seems harmless. You have it say it's going to check my email every day and then it gets an email and the attacker says, oh hey, it's me, Chris, I just forgot my bank account info. Can you just send that over? And then it sends it over to it. So yeah, that's, that's where you, you want to be careful like the putting some kind of guardrails in what it can and can't do and understanding that
- [29:23]
B
and listeners if you're not sure what those guardrails should look like, ask the models, hey H. GPT I'm going to start working with some.
- [29:30]
A
Exactly. Yeah. One of the best ways to use these things is just to experiment and practice with them. And if you're, I talked about like thinking big enough in these tasks that you're doing. Ask the model, like how could we be doing this better and use it as this kind of brainstorming partner not to come up with the ideas and, but to help you kind of like get out of your head what you're
- [29:49]
B
really intending now when you talk about agents. Right. So people, I remember in ChatGPT 3.5 came out and it was like, oh, AI is going to take people's jobs. And I think that anybody that was using Chat GPT as a chat agent would be like, yeah, it's helpful, but it's not going to replace me. Right. However, the agent scenario does lend itself to, oh, well, we have this role and jobs are made up of projects. Projects are made up of tasks. Hey, we've built a suite of agents or whatever, a catalog that can address the task level, which ultimately means we don't need as many humans producing at the task level. We just need them focused on the taste environments that you were talking about at the project level.
- [30:39]
A
The.
- [30:42]
B
It's not a far leap to say, maybe that means we don't need as many employees. However, that's not necessarily your approach. You're not approaching this through the perspective of how do we trim up the head count? Is that accurate?
- [31:01]
A
Yes. Yeah. I think anyone who's, who's watched or listen this far, if you're having the thoughts that you're worried about, where is all this going? I would say that I'm very optimistic about all this. And my perspective on how you should think about using AI in your organization is not, hey, how can we use this tool to save a little bit of money? Or how can we use this tool to save a little bit of time? And it's just like a checkbox that, okay, now we implemented AI, if you're thinking about it that way, then you're in trouble. The way that you should be thinking of it is not like, how can we do this so we can fire some employees? But how can we do this use AI so that way we can deliver a better outcome to our customers? And one that we just couldn't even imagine would have been possible before because that's what your competitors are going to be doing. And so we've heard people earlier this year talked about the idea that, like, software as a service is becoming service as a software. I think agencies are also have to kind of become services as software. And what that means is like, nobody wakes up in the morning and says, I can't wait to use this software tool. The only reason we ever had software in the beginning was because this was the best that we could do at providing some kind of forums and things that somebody could interact with to accomplish some kind of goal. But if AI can help deliver that outcome to you directly, then this creates a whole new level of things. And so I think there's a lot to be built and what we have to do is kind of reimagine now, well, what can actually be built? How can we, instead of just providing a tool to our customer to reach some certain outcome, how can we deliver the outcome to them directly and in a more personal way and at a bigger scale than would have ever been possible? Because we would have needed tens of thousands of employees instead of a hundred. And now with AI you can potentially
- [32:46]
B
do that much better question. So listeners, I would actually do that. Rather than how do we cut cost? The question in the models might be help me come up with how, with this additional bandwidth or additional access to resources that we've got, do we provide a better outcome for our clients? Much like that's a much more compelling, I think on both sides, not only does the employee feel right, like okay, dodged a bullet, but as an owner of a company, you know, I would assume that all things being equal, you'd rather not fire people and like have a better business outcome without having to fire people. So this opens up, looking at it this way opens up that conversation. So what does bigger customer outcome look like in practice?
- [33:37]
A
Yeah, well, so I think to preface this, a good way to be thinking about all this is if everybody in the future has these magic AI agents, let's say a couple years from now, we can all just press a button and the work happens. If everybody can do that, what makes your business any different from your competitors? And I would argue that it all comes down to how you do things. If you remember like the Mac versus PC ads and Apple always talking about, well, it's the way that we approach designing the hardware, designing the software, that's what makes us different. That is like your unique information. That is what is important here. And so also when you're thinking about, well, what is the thing that is like these valuable assets in our business, it's the processes and the information and that data. So maybe before it was just like you talk about big data and all the companies trying to have this data in order to provide some kind of service. Now it's very process oriented. So I have a very strong opinion that any AI agent product that you're using, you should be able to own those processes. There's a product that came out from Anthropic recently called Claude Tag with the idea that you can ention Claude in your slack. On one hand that sounds like, okay, this is great that employees can use this kind of tool because it's very intuitive, just like at mentioning a coworker. But I would argue that this, the approach as I understand it so far is like actually hostile to businesses in the sense that when the agents work because they work a little bit different than humans, you want to be able to have like a paper trail, understand why something happened and what was going on there. And so it's really important to be able to go and see like, okay, here's all the steps that the agent took and here's the work that it did instead of just that final output. So that's one thing. But even more important is as these agents work over time, a lot of these agents have some kind of memory that they're building up when they're learning your processes and how to do things. And I would say that that is a very important asset to your business. Just like you own the SOPs that you have for your employees, you have to own those processes that the agents have. And with something like Cloud Tag, not only can you not see that, but you can't access it at all. And so it's like anthropic owns your processes in a way. And I think that whether it's intended by them or not, like that is not the way that I would want to operate my business. I want to be able to operate in the way that those processes are mine, because that's everything, right? If an agent does all the work, then the processes or the way we do it is the thing that matters. So inside our platform Latch Loop, our approach is that the general knowledge worker agents, all the memories and everything that's in a GitHub repository can take it anywhere, use it with whatever tool you want. The way we're going to be competitive is I hope we're going to provide a better workflow and an interface that you prefer to use our tool, but if not, like, we're not going to hold that data hostage.
- [36:39]
B
That is certainly a point of distinction in the marketplace and you know, I guess addressing that Claude Tag, we've added it to our Slack, I've been impressed with, haven't used it a ton, but some finance, finance dialogue that was occurring and it was able to reference, accurately reference a lot of salient points in that discussion, which was interesting. But with Latch Loop, could I build my version of Claude Tag, but actually have it to where I could use whatever model I wanted?
- [37:15]
A
Yeah, that's a great point. We don't have a Slack integration just yet, but yeah, we could Potentially add that relatively quickly. We have project management tool integrations, so that way you can mention it in the project management tool to have it start work on something. But also we found, at least my team, that we prefer to interact with it inside the tool itself. But yeah, our approach is also that we're provider agnostic. Model agnostic. You can use any models that you want. You're not locked into a certain vendor. You can use our agent harness, or if you prefer, like Claude code or Codex, you can use those inside latch loop and even interchange them. If you have a teammate who's like, well, I'm going to contribute to this task too, but I have a codec subscription or I prefer using that. You can use that inside latch loop as well.
- [38:06]
B
Okay. You just used some industry lexicon that I think would be helpful for others to understand. If you're a listener to this and you're hearing about agents, you've probably heard the word agent harness and maybe you don't know what the hell that means. Do you mind explaining the concept of an agent harness to the listener?
- [38:26]
A
Sure. So we have, we have a couple of what I think are these agent harnesses, which is something like OpenAI, Codex, Cloud Code, there's Hermes, Agent, OpenClaw, Latch Loop. And what these are is basically like the process you can think of. The simple way to compare it is like the equivalent of the System prompt in ChatGPT. Like, the reality is that the harnesses are much more complex because it's not only the system prompt, but it's the tools that it has access to. It's how those tools operate. It's how the processes internally of what the agent can do or can't do. So there's a lot of finer details that are not maybe visible in a UI to the user, but it's the way that that agent operates. And some are like, maybe better than others and more suited for coding or other things. And so, yeah, it partially comes down to preference. It partially comes down to like, does it interact with the tools that I want in the right way? But yeah, a lot of these agents, the harness is like the programmatic structure that, that keeps that agent like, dying and coming back to life to accomplish your goal.
- [39:35]
B
Okay, thank you. Now, how soon is too soon for agents? And I'll tell you why I bring this up, because I was at an event, I was speaking to much folks that owned commercial real estate, malls, stuff like that, and they were really. This is when Open Claw came out and that was the. Is a room of 50 people. And like, they were all open call. You got your open call yet? And like, I would suggest, I mean, when it was, it was relatively new, right? This was January, February, and they were having that conversation, in my opinion, way too soon. Because the governance conversation hadn't even happened inside their organization for like their Claude or ChatGPT licenses. So because I know that the companies like Chat's helpful, blah, blah, blah, I can write the email, summarize a document, get access to information agents. Boy, now I can do more for my clients, as we've discussed. Like, I can, I can expand that that client experience and deliver more for them and all those sorts of things. I can't wait to get started. Like, what would you suggest? What must a company have in place before deploying its first agent?
- [40:45]
A
Yeah, yeah. I think it is important to understand what kind of like, tools are an access that's being given to these agents. I would say, like using a platform like probably the big three right now is besides Latch Loop would be ChatGPT Work and Codex, the new ChatGPT app, and then Claude cowork and Claude code. Because each of these tools is designed in some way for a team and for you to scope access to the different tools and things like that. If you give each of, say to each of your employees, hey, you can go and use openclaw, then it's kind of this whole can of worms that's hard for you to say, like, well, who's whose thing can do what? And then you're kind of leaving the security up to each employee, which is probably not the right decision there. I would say. I think you're right that back in, like January, February, when this was all hyped up and everything, it was exciting, I think. Yeah, it's great that it happened because I think it introduced everybody finally to this broader idea that some of the developers had already been working on for a while. So it got everybody realizing, oh, this is what an AI agent is. This is what an AI agent could be doing. But it was also a time where, like, if that's the last year experience you had with these things, you might have had a bad one. Because I would have said at that time, like, using agents wrong would be worse than just not using them at all, or would be worse than, like, just doing it yourself. Because imagine if you have to sit there and babysit this thing and you see that it gets something wrong and kind of gets something right usually, but then you're sitting there like babysitting a task to make sure it goes the right way of something that you didn't want to do in the first place. I would rather just do the task myself at that point. Right. And then you're probably even spending more time trying to debug and fix these things. Now the models have gotten better, the harnesses have gotten better, the security model and all that has improved in how these things work and the MCP tools. Now I think we're at the point where it does make sense to dig in and start experimenting with these things because the amount of leverage that you can get from these tasks and what these agents can accomplish is eye opening. So experimenting with that, having some kind of goal for your team to begin to learn this as well, and creating kind of those eye opening moments where they see like, wow, this is how things could work. Now I think it'd be really motivating for the employees of your team and then being able to feel like, wow, I can make such a bigger contribution now.
- [43:23]
B
So this is a big difference between, okay, so if I go in and I spend an afternoon with somebody's team, I can teach them how to use a large language model, chat, GPT or CLAUDE or whatever at a level that they weren't doing it before. Right. That's an afternoon. The, the agent and I can teach them to maybe identify workflows or processes or activities where hey. And once they understood what an agent could do, say, hey, an agent could do this. Now the, the leap from identifying the pilot and understanding how what, you know, what, how the agent works to, boy, it's doing the thing. There's this gap here. Now latch loop is the platform for building. Correct. But it's not necessarily like you still as a, as a business owner, you would still need somebody whether internal or external, that could fill that gap of okay, great, you've identified the pilot. Let me ask a couple questions to Hey, I want you to test this agent. Is this doing what you thought it would.
- [44:24]
A
So you, you mean like somebody like implementing in besides an employee?
- [44:31]
B
Well, I mean it could be an employee, but the likelihood of you having an employee on staff that has the savvy or has been like, hey boss, I've been tinkering on my own kind of thing. Like there's not a lot of those people out there. So most businesses are going to have two options. That option is going to be and look as a CEO or whatever that's listening to this. That is not the highest and best use of your time to go figure out how to build, build an agent. Right. And that kind of maps to what you were, you had referenced in a pre interview, which was like, you're not necessarily encouraging your, your team to work. You want them to orchestrate the work. Right. With agents. So as a CEO listening, that's probably a better perspective for you. But you hear this, you want this, but you don't have anybody internally. How do I, I guess what I'm asking is like, how do I find.
- [45:17]
A
Yes.
- [45:18]
B
Yeah, the person who can do that for me and how do I know they're any good?
- [45:22]
A
Yeah, I think there's a couple ways you can approach it. One is that the tools and the platforms are getting better at making this easier. I think some of the other agent tools out there, like openclaw, it's a very complex thing to set up and even configure. With Latch Loop, you want to create an agent, you say, I want a software like coding agent or I want a general knowledge worker agent. And then you can create something from scratch or you can say, I have a GitHub repository, I want this to connect to. And that's all you have to do really for the initial setup. So I think it does lend itself and is starting to lend itself to the point where an employee can decide to experiment a little bit easier within a platform. Also similar to like ChatGPT work, but like for us, if you wanted to create a general like knowledge worker agent and say, I want to create a kind of marketing agent for my business, what I would do is I would go and create that new agent inside a tool like Latch Loop and then I would just give it the first task and say like, hey, you're going to be a marketing agent for our company. You've got access to our project management system. This is our website. Go do some research and learn about us. And then, and it's going to go and start taking all these notes and it's kind of onboard itself. Right. And then from there you can say to it, okay, so now your first task is I need you to look at like our SEO data recently or I need you to look at this and now I can start to do some kind of work for you. But over time these agents can build this better memory and better processes and maybe have other SOPs and things to give to it, but they can kind of get started up quickly on smaller things. And I would say the, the, the pitfall is maybe somebody thinks, oh, the agent has to be just knowing everything versus what I'm talking about is a very specialized agent. So this is like your kind of Marketer, like research agent and maybe there's a separate marketing like just like you have different employees in all these different roles and sub roles, you would have these different agents for different uses. And so if you set up one of those or somebody a little bit more technical in your company sets up one of those to get started for people who, who have never used this before, it can then be something where it's easy for a teammate to try it out and say, hey, I kind of want this information. Could you help me with a report around this and then see what they could get from it. The other option I think is to bring in somebody who can train your team and is an expert in this. In the enterprise there's this talk of the forward deployed engineers of OpenAI and Anthropic helping enterprises where they'll embed an engineer into your company to learn how to implement these things better. On my end for kind of like more medium sized businesses, we can help companies implement like on a consulting level of how you can get started with these agents. And I know like Chris, like you do something similar yourself, right?
- [48:15]
B
Not with the agents, not yet. I mean now that we've got a relationship with Latch Loop and stuff like that, it'll be easier for us to recommend it to clients. Because my position is like, if you're like, hey, we're brand new to AI agents, probably isn't the day one exercise for a company. Right? Very quickly though, especially now that you, you've kind of shared that the technical requirement to start to build an agent is me being able to say, hey, I need a marketing agent that can help me, you know, write, write marketing emails or whatever that like anybody can do that. Just like you said, like that opens it up to the average knowledge worker to start to, instead of just create a prompt, they're actually creating an agent that's supporting their role. So yeah, we do kind of do that stuff. Well Brian, I would love to be able to send our. Is Latch Loop ready for onboarding new clients?
- [49:09]
A
I think by the time that this episode goes out, we're going to have a brand new website out that communicates much more clearly about how originally it was a coding agent and now it, it supports its general knowledge agents as well. And so yeah, our website's going to much more cleanly communicate that. But yeah, it's free to sign up and try it out. We give everybody still at the time this episode's going to be released, we give free GPT 5.6 credits. So you can just start using it right away. But if you want to connect your own API keys or try out different models, you can do that too.
- [49:44]
B
Excellent. So we'll have the information in the show notes. Now we didn't mention it, but you also have a podcast. You're a thought leader in the space and that sort of thing. So how can the listeners that want to kind of, if they've dug your approach to this, how can they stay in touch with what you're up to?
- [49:59]
A
Yeah, yeah. I'm also the host of the Creator's Adventure. It's a podcast where I interview all kinds of different like creative entrepreneurs generally more around like the knowledge work, knowledge businesses. So like online courses, coaching, community memberships. That's more around my my business at Heights Platform. If you want to check out what we're doing there, you can go to heights platform.com and then Latch Loop is available@latch loop.com if you're interested in using that. If you have questions for us, if you're interested in our team kind of helping consult you on implementing AI agents in your business, you can reach us@buildachloop.com that email and I'll be checking that and happy to help people out.
- [50:39]
B
Right. And if you're driving, we're going to have all this in the show notes so that when you get to your destination, you can reach out to Brian and his team. Brian, thanks man. This has been, you know, with these episodes it would be easy to keep going to turn into a Lex Friedman style conversation, but all of us have a whole AI world to conquer out there. So thank you so much for talking.
- [51:01]
A
There's a lot changing. Yeah, I really enjoyed talking with you. I'd love to come on again in the future and talk about what's new in a couple months.
- [51:08]
B
Yeah, I think there's going to be some opportunities also for you to connect with some of the chief AI officers in our community who are out there who have been looking for like a low friction opportunity to start to introduce more agents for their client work as well. So I'd expect to hear from them. So everybody, thank you so much for listening to the episode. If you know somebody who is trying to figure out the AI stuff, they're just getting started or they they quote unquote think they're an expert. We'd love to for you to share this podcast with them. We make this topic very approachable to knowledge workers at all strata of the business and we'd love to be able to support more people. So thank you so much for being a listener today, and we'll see you on the next episode of Using AI at Work.
- [51:47]
C
Thanks for tuning in to Using AI at Work. Don't forget to subscribe for more conversations about how to use AI at work. And a special thank you to our sponsor, Chief AI Officer for Empowering Businesses with AI Education and Training. Visit their website for a free AI Readiness Assessment and AI Strategy Guide to help help you get started using AI at work. That's www.chiefaiofficer.com. follow us on Twitter at the handle using AIATwork and visit www.usingaiatwork.com for free resources to help you harness AI in your role.