58: Using AI for Building Multi-Agent Systems with Jake George

This podcast episode is sponsored by chiefaiofficer.com offering training and certification through the International association of Chief AI Officers. Interested in a new career or leveling up your value in the marketplace? Chiefaiofficer.com can help. Welcome to Using AI at Work. I'm your host, Chris Daigle. 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. Hi, everybody. Welcome to the podcast today. This is going to be an exciting episode because of the topic. This is a hot topic in 2025, obviously. AI agents, where they're going, what's possible with them. And the best person that we could have talking to us is our guest today, Jake George from Agentic Brain. Uh, so, Jake, before we get started, if you don't mind, just kind of share with everybody. In particular, what I'm looking for is your journey to becoming this agentic expert.

Yeah, so where I started is. Oh, boy. A couple of years ago now, it was like, right when GPT 3.5 came out. So what I was doing before then is I would build and grow developer teams. So we were mostly working in the crypto space. We were with this crypto marketing agency and consultancy. And so I ran their whole development team because oftentimes what they would have is clients that they were consulting for would need smart contracts written, which is essentially like the software that makes the coin or NFT for anyone that doesn't know. So right about that time, GPT 3.5 came out. So of course, like, everyone, you know, I hop on, start playing around with it. It was just like this light bulb in my head. Like, I. I'd known AI had existed. I had used it, like, briefly before. But as you know, before then, it wasn't really that good. There wasn't so much, like, chat with it. I do some of, like, the, like, Google models that would do some of, like, the, the video and image rendering. And you know, back then they were like, really, like, not that impressive. So it was like I had kind of heard of it and I was just like. But then you get on GPT 3.5 and it was just like a light bulb in my head. And I was like, oh, my gosh, like, this thing is really legitimately smart. Like, it can figure things out. And so then I'm like, okay, sweet. It has an API, like, what can we do with it? And so, you know, it just starts out with just like, oh, you know, write this for Me create a marketing plan, blah, blah, blah, like write this little bit of code. Where's the bug in here? But what we started doing was building out automations for the internal processes and the sales processes for that company and then seeing the, the results that it had for them. I then broke off and then created my own company, which was the one that I ran before this one where we did like workflows, AI automation, and even like at the point of GPT 3.5, agents were not really a huge thing. Although technically, like possible. There wasn't as much like tool calling code running, all that sort of stuff. So we mostly focused on like automated workflows. And then once agents started to become a thing with like, you know, like auto gen relevance, what else do we have? Like LangChain, which, I mean LangChain existed before then, but their agent abilities got better within. So then after, you know, I just always have been up to date on what's coming out. So then I saw agents and I was like, wow, this is the next big thing. Because the cool thing about it is you don't really have to plan out every step in the workflow. It can be completely dynamic and you're like, all right, here's five tools and here's our end goal. Just do the thing. And so if you think about it that way, almost sort of like builds its own workflow dynamically and intelligently. So that's where I really saw it, where I'm like, it's useful before, but, but when it's able to call specific tools on its own, reflect on what it's doing, have memory and make decisions based on the information that it's getting while performing a task. That's when it really gets to the point of like this can actually act as a sort of like low level employee. And then obviously the models get better, the tools get better, it gets smarter. We have reasoning models now which are not, you know, not so useful in like agents yet. But like they will, you know, they will be. So yeah, that's, that's really how I got started. And then I met my business partner who was unable to be here today, unfortunately. Frank. And one of the biggest things with my clients is that they would, when I was doing Synthoria Labs, my last company, is that they would have these very broken internal processes. So like, we want to automate our sales process. Sales was the main thing, but just seeing how it worked, okay, like what's your close rate? And it's like 5%. It's like, you know, yeah, you can automate it, but that doesn't mean make you're not going to go to 20% close rate because you automated this like AI does not close for you, it doesn't sell for you. If you have a good process, it can help you and eliminate that extra like manual work that's like, can give you back more time to sell. But there was like a misconception that it would sell for them. So meeting my business partner Frank, who's been a fractional CRO for the past 17 years, he just is immensely skilled at being able to like analyze and break down these sales processes and being like, yeah, we'll implement AI but also adding a consistent sales consulting side to it to be like we can add AI, but these sorts of things need to change which we can help you change and automate. And then also with like a little bit of coaching and consulting can really actually improve their, their sales. And that's really at the end of it. Like companies love to like save time with AI, have convenience. But when you can help them improve sales, that's when it really clicks and they're like okay, I, I love this. Like that's the main thing that gets people that, you know, makes people believers essentially.

It's a great tip for anybody who's listening that is interested in being like an AI consultant or anything. Sure they want help with you know, repetitive tasks and things like that. But our recommendation is very similar, Jake. It's like start where they can see an ROI impact and then the rest of the stuff becomes easier to get a yes from. But if you can impact sales, they get very enthusiastic about AI. One of the, I think, I don't know if it's a misconception or misuse of the term for sure that exists is, you know, I've been involved in generative AI as long as you have and agents aren't a new word but I think the, the definition or the expectation of what an agent is. In the past I've seen people that were building automations and calling those agents. What is your definition of an agent? I mean obviously limited to the capabilities we have now, but what would be your definition of it based on difference between an automation and an agent?

Yeah, so I would say the, the main difference is being able to like dynamically make decisions and call different tools without having a sort of planned out workflow. So it's like sort of the prompt then in a way is their workflow. So it's like agents are able to, they're, they have memory, they're able to reflect their goal oriented and have tool usage. So these are all things that humans have. So if we go back to even before AI there were automated workflows, you know, those that existed for a long time. But you kind of have to think out every sort of step of like let's say you're using makers app here or something like that. You sort of like think out the path that it should go through and it can work really well and especially when you're using. Because we do a lot of custom software, you can get a lot more in depth and help with the sort of like decision making through the workflow. But the main thing is sort of just giving AI a goal and it being able to like extrapolate. Okay. If you're like a check on this lead and see where they are in the pipeline and then make the next step. That's pretty vague but it's like you could tell a human that and if they're trained on the job they'd be like okay, well what should I do? Probably check the crl. Okay, let me go find the lead. Okay. It looks like they were supposed to have a, they're on vacation, they're going to be back next week. The next step would be scheduling a call. Well, let me see. It looks like the call hasn't been scheduled. Okay. Therefore I should schedule a call with them. How should I do that? I should email them, reach out, say hey, hope you had a great vacation based on the context of what went on with the lead. Do you, you know, do. Let's, let's have that, let's have that follow up call. So you could say that to a human but that would be very hard. Like in a workflow you could make a workflow for that, but you'd have to anticipate that, you know, that chain of events happening every time so that there's a way for it to go. Or it might just be very templated of just like, you know, you could do like follow up with this lead and it can send like a templated email just like hey, just wanted to see where we are. But it's like kind of being the goal oriented, deciding what tools to use and then being able to like dynamically shift its path based on the information that it's receiving back from them. So like what if it said check on the lead, they already had a call schedule then therefore do nothing because that was the next step. It already happened. So that's kind of. And then being able to reflect on what it has done. Go out and find information and then change its course of action based on that information that it has. So that in a very, in a nutshell is what I would consider an agent versus a workflow.

So any business owner or team leader or any, anything like that that was listening to that, they might hear that sounds like a human, right? And you're saying that that capability exists currently with the, in the right hands to, to build it out obviously. But businesses could have that type of decision making and dynamic interpretation of next course, you know, next step in the, in the activity automatically decided without the human necessarily needing to say okay, now do this.

Yes, yes. And it's there, it's kind of the same thing because there are not the same thing, but there's, there's two parts to it. There's like tools, SL workflows and then agents. So a lot of times what we'll do is because a lot of like agents will often call automated workflows to complete actions or just call straight tools. Okay, so a lot of times what we'll do, like to lay the framework of like what we do for our clients is we would start off with like usually AI automations to sort of be like, okay, let's make an AI automation that takes the context of a lead and sends an email. You know, just writes an email in Gmail and saves it. Okay, that's a sort of automated/AI workflow. But then really what we do is like as we add more and more, we start adding them to like a, we'll have like a manager agent and then like a bunch of sub agents and usually we'll have them communicate through Slack. It's a very good channel for that. A lot of people use it and you can have your whole team in a group chat and then you can just ping the manager agent when you want something done. So the idea is that we sort of start automating all of these parts here that they can use immediately. But then as we continue we sort of tie them into more agentic workflows. So it's like this one, this one might manage the CRM and be able to make actions in the CRM. That's cool. It's an AI workflow and you know, it might be an agent itself, but the cool thing is when they can all interact and then the manager agent is like, okay, well we got to find out what's going on with the lead. Let's query the CRM agent. Okay, cool, now I got the information back. Now let's query the email agent okay, now they said that they're free on Tuesday at 4pm now let's query the calendar agent so you can have like AI and automated workflows are the base of it. Then those go up to different agents. And it's better for agents to focus on like a very small amount of specific tasks because although there are some like Google models where they have context windows of a million, 2 million tokens in action, that doesn't, that doesn't mean that you can fill it with that much information and it's accurate. It is proven actually that it's very inaccurate and we can even see that just from testing without doing a million tokens. So then you have like your, your worker agents and then those are best to go to a manager agent where the main job is it's pretty simple, it's just like what's the task? Break it down. Which agents do you tell to do that? And then they go and do the stuff it queries all the agents, they go and do the actual work, let's say, and then send the information back up. So then at the end like there's not really, and I think a misconception that a lot of people have is that you're going to go, you're going to have your own specific company needs and how it operates and everything and you're just going to go online and buy for $20 a month the business running agent that's just going to do everything for you, figure out everything and grow your business. And it's like that doesn't exist today. And I don't think it's going to exist for a while. So it's like, it's really built on workflows, automations and best practices data sub agents and then a manager agent. And that's really where you can actually see the benefits of like oh, okay, so this, it does, it can exist but you're not just going to go online and click download business running agent. It's just, it's not chat GPT, it's too, too in depth for that.

So for anybody listening, one of the things that helped me kind of understand what Jake just shared was it the agentic architecture looked like an org chart. And everybody knows what an org chart looks like. There's a level of decision making, then there's a level of action. There might be a level of action below that with each individual on your org chart having a role. The agentic architecture that I've seen for some of those, it looks exactly like that. Like you mentioned the manager Agent the. Do you. Is there a specific lexicon for the different layers of the decision making?

Probably I just call them the manager agent, sub agent. That's just kind of the terminology that we go with. I've heard that one commonly used or worker agents. So yes, if you probably, if you look it up, there is a technical definition, official language. Yeah, I just kind of, I always relate things back to business because it's like. And that's another thing with like working with clients in the business aspect of stuff. It's like, you know, to really see the effect. It's like you need to have like the technical knowledge and then like the business knowledge as well to see good effects. So I'm always relating things of like sort of translating AI terms into like how does this make sense and how can someone picture it in their head? Because if they can't, you know, if I use really technical jargon and stuff like that, they're gonna like, I still don't understand it. I don't see how it helps. But if you're like think if you have a manager and then the manager has all these employee agents and that's what you're gonna get. And then they're like, oh, that makes sense to me. So that's what I call it.

So the activities that are occurring at the manager level, the sub agent and then I guess even a sub. Sub agent, what is the difference between how they're behaving at each level?

Yeah, so the first one it would be like, let's say that you're in Slack, okay. And you just want something to be done. Let's just use the same example, the check on this lead and make the next steps. Okay. So you would then through Slack you'd be querying the manager agent and then from there the manager agent would decide of like, you know, because you explained to it, you explain to the sub agents what tools they have and what actions they can take and how they work. You explain to the manager agent what the sub agents can do and what tools they have and what goals they would accomplish. But at a simpler level because you can just say Calendar agent takes care of anything calendar and scheduling related. You don't have to explain to it it has access to Gmail or Google Calendar, create event like it. It doesn't really matter. So it's kind of you just break it up like email agent controls the email CRM agent controls HubSpot, calendar agent controls calendar that you know, and, and kind of just break it down there and then just have it Sort of come up with a plan and then be like, your job is to then query these agents.

So.

So based on the task coming in from the user and your knowledge base, which ones are you going to query? And then it kind of has to give them a short like hey, HubSpot agent, check for this lead and see what the next step is for them. Okay, here's the next step. Okay, Gmail agent, write them an email based on this context. So that's how it would work on the hierarchy of things. Another thing we're experimenting with right now and seeing pretty good results because in, because right now we're not using O3 mini for like the actual agents is it's not always necessary. I think that it would be very good for that. But right now it's more of like just for reasoning and not so much for like tool calling but doing a step before. So query comes in and essentially treat O3 is like the manager slash planning agent because it reasons very well and then bump that query over to the manager agent which then it kind of just relieves that one which is, which we would do on 4 oh, you could even do it on 4 oh mini. Because if O3 does the reasoning passes it to 4 O, all it's doing then is tool calling to the other ones. So that's an interesting way to incorporate reasoning into it. For most tests I would say it's not absolutely necessary as 4O is pretty smart to build out plans, but it's just, hey, it's half the cost. Okay, it's very fast and it actually reasons so you know, you might as well and you get really good results. So like planning and reasoning is like that's pretty big for the manager agent. And then when that gives instructions, you can have it also write structures like so what should be done? And then what should be the instructions to each sub agent and then it can just pass that on.

So for this manager. So first off, one of the things that we do in our Chief AI Officer certification that people really love is we dive into process mining and process mapping. Obviously the steps of a process for automation at least need to be very clearly defined so that that automation can deliver high integrity, consistent output as if a human would, if not better. So beyond the process idea, let's say, let's say it was a role for an HR coordinator. That's something, you know, a lot of companies have the HR person, when you're doing this, do you need to know what the HR like, do you need to have a process map or to be able to do. So you guys work. That's a big part of what you build.

Yeah, so what we do when, when we're working with clients, we always mix like consulting and AI because it's just like if you're not understanding their actual problems and workflows, then there's, you know, you're just making guesses. So mostly we focus on the sales side of stuff, but that can also touch like client success, marketing, HR somewhat. So that's, that's really where we, we spend our focus. But yeah, so what we do as you know, is like our, our sort of scoping call where we sit down and we actually take like 90 minutes to like, okay, what, like, what's the biggest issue to tackle right now? And like, how does that process actually work? And so then we would dive into sales process of like, you know, how many SDRs do you have? Are they making calls? Are they following up? Are they doing pre call research? Or you know, are they, are they actually logging this to the CRM? Are they using it properly and then seeing where kind of like, where's the bottleneck, where's the inefficiency and what could be more automated and done better? And then from there we have that whole process and then we can analyze from there either like, is this process actually good and where could it be better with a few small tweaks? And then where should AI be added and where should it not be added? Because a lot of people come like, well, I want AI to do cold calls and close deals for me. And it's just like, me too. That's. Yeah, it doesn't everyone like? Yes, I. So we joke about it. We're like, yes, we have the Infinite Money printer AI and we would just love to distribute it to everyone for a very low cost because why not? Why would we not just use it for ourselves?

Yeah, this is interesting because I encountered the same thing. We talked to clients on the service side of our business and they're interested in AI enhancing their, their operations. But before you can even start with that, you have to make sure that the process isn't wonky. Like you referenced that, you know, the 5% close rate example earlier. That's like, fix something else before you introduce AI because there's something missing there. So I think that that's a big takeaway for any of the listeners that are excited about deploying AI. You may have to do some groundwork that has nothing to do or not AI, but agents, you may have to do some groundwork that has More to do with like how you guys are operating your business than it does with anything to do with AI. That's a great point. So listeners, there's no miracle you might have to do a little work here to get this agent to work.

Yeah, there's. There seems to be a misconception unfortunately. Some people see AI as a way of like, oh, I want it, I want it so I don't have to think and I can be lazy. And it's just like, you know what, no matter how good AI gets, you're just not going to be that successful in business if that's your whole idea. It's like, you know, it's more for like oh, I want to like think of higher level tasks and don't want to spend my time on say customer service call where I'm not really thinking too much and it's just a waste and repetitive and whatnot. A lot of times they see it as like, oh, I don't actually like know, know how to run this part of my business. I don't want to figure it out. I don't want to have, you know, do trial and error and figure out a better sales system. I just want AI to do it for me so I don't have to think. And it's just like that's using it the complete wrong way. If you're using it to be lazy or not figure something out A you're not going to get a good result from it, it's not going to figure it out. It might try and you would have a very mediocre or poor result at best and you've really made no progress. So yeah, we try to avoid people with that sort of mindset and look at people who are like, you know, experienced serious business people that are more like, I want it to do repetitive tasks that don't take a lot of thinking so I can do higher level work. That's kind of the idea.

So if, if anybody pays attention to any of the social channels that have AI as as a topic, whether it's Instagram, TikTok X doesn't matter. We see a lot of people making posts about like oh, this agent. Right. They're loosely using that term is is what I'm seeing on those social channels where somebody's got, you know, four or five activities or something. It's not terribly sophisticated. They're building it N8N or make or probably more like N8N is the ones I've seen that have had I guess the most robust capabilities. That's not what you're talking about though.

I mean, what do you mean agents versus workflows again or what?

The kind of things that like if a client says hey, we need help with our sales process, we're getting good results on the process side, we just want to be able to scale more or scale with fewer people. Is that's not the kind of thing that they can necessarily find on a five minute TikTok video?

Yeah, no, I mean technically for those. Because there's, there's a lot of misconception, there's people that will confuse like AI workflows and agents. But then there are also sometimes people would be like, well that's a workflow. But it's like agents don't always have to be super complex and technically will still satisfy the definition of agent because it's using AI and intelligently calling tools like anything. Like you know, the OpenAI Assistance API, you give that a tool or two. Like technically that's agentic because it's dynamically using AI. So it's like there, you know, I can't really come down too hard and like, oh, those aren't agents, they're not agents. It's like, yeah, no, technically they are. If it's using AI to intelligently and dynamically call different tools and make decisions that technically suffices as an agent. So. But yes, they are not on the level. Like are they going to be incredibly useful? Probably not so much to be honest. If you, if it can be, if, if you see someone make it in five minutes or less, I mean it's probably going to be a very simple agent which not saying there's no merit because like I was explaining a lot of times you need to start with the simple and fundamentals. Like you know, you, a lot of times the clients, it's like they need, part of their, part of their sales process is sending emails. They need to have an email agent that can understand context and dynamically decide whether to read emails, reply to emails, draft emails. So even though it might just be AI using Gmail, it still suffices. You know, that's an agent. Okay. But what makes it more useful is when you have the whole team of agents which is very complex and you will not see a lot of that in simple Instagram TikTok videos. Usually they're looking at a very basic agent where it's like this one writes my social media content and it's like cool. But literally you can do that with a fine tuned GPT. Like that doesn't, that doesn't need to be agentic? It could be, but it doesn't really need to be. And nor is it that like amazing and impressive. So really the combination of workflows, AI automation, agents all together and then tied to one agent is where you actually see really good results.

So for everybody listening, I think that we've just busted the myth that there is the hey AI, go and handle all of my marketing. Like that. That agentic build is not for sale at this point. So it's interesting. Then tell me if I'm on the right track. I, if I'm seeing something like maybe this guy's got one about writing social posts and this one's got something about, you know, creating, I don't know, a, a video through one of the tools and that sort of thing. Individually, they're kind of, oh, that's fun, that's interesting. Not a lot of impact. But what you're suggesting in the architecture that you guys do is that let's say it was a marketing agent, that there might be like dozens or 10 or whatever of those. Here's my sub agent that does post. Here's my sub agent that does the images in ideogram. Here's the. But there's still that coordinator level that is taking the input from the sub agents and it's making a determination on how it either assembles or what it uses or it's making like an additional layer of judgment. Almost like Human in the Loop would.

Yeah, yeah, essentially. I mean, it's kind of like a agent in the loop. If you can say you can, you can still integrate Human in the Loop to those as well because you can have it sort of make it as, you know, like, do part of this process. Check with me on the results from the sub agents, then continue. So that's a cool thing with the prompting too. You don't have to readjust the workflow. You can just be like your prompt is essentially your, your dynamic workflow. But, but yes, you got it exactly right. Of like a lot of times people, sometimes people will portray these complex multi agents in a very simple way just for like, you know, views attention like it's part of marketing. Sure. Like, I'm sure you've seen it the biggest. Pretty much anyone puts out a YouTube video titled I made an AI agent to do anything and you're guaranteed like 100 to 500,000 views. And it's just like. And then what it actually does is it sends messages in telegram and Gmail and it's just like cool. Like that's not everything and it's nor is it that useful. It's just like it's certainly cool but it's a such a clickbaity thing. However a lot of those like if you take that concept, oh, the AI agent to do everything. So I think a lot of people like oh, that's what I've been looking for. I just need one to do everything. But really it's this whole team of tools, workflows, AI automation and then sub agents and then like a main agent. So technically yes, it is an AI that does whatever you make it to do which is not everything at this point. But I mean hey, good for views. But yeah, that's, that's kind of the idea. If you see anything that's really complex, I would say 99% it's a multi agent team working together.

Okay. If you're enjoying this episode and want to learn more about how to start using AI at work, we've made it easy for you. For just $1 you can have full access to the Chief AI Officer community which will give you additional training, custom software, daily Training calls on AI tools, using AI automations, getting more from your ChatGPT sessions and the business of being an AI consultant. Simply go to chiefai officer.com insiders to accelerate your AI journey. Now back to the episode. So these guys, because I've seen those and it's like it was this, you see the diagram and you're kind of like, you're like oh my gosh, that looks. But they're like watch. I can talk to Telegram, it can check my email and tell me that there's an E. Like even though it's very sophisticated looking, the application that they're using it for isn't. It's basic things that might certainly didn't need to be over engineered to that level. So it seems like some of these experts out there, they've got the idea down but they don't have the application down necessarily. Like a business doesn't care if they can talk to their Telegram channel and find out if they got an email and who it's from. But they do care if it's something like what you talked about earlier with the that sales environment or that's setting the appointment. Well, I'm checking the notes from the contact record, it's saying this, let me go to the calendar and book it for when they're on. Okay, this is helpful. So as you're obviously paying attention to what's going on in the, the agent chatter that's going on in the space. Where do you think people are having the biggest, the wrong assumptions about? Is it, is it that, that I can do something and it's everything does everything for me or what, what are kind of some of the things that you see people making mistakes with when they're talking or thinking about using agents in their business?

Yeah, so I mean there's, there's definitely quite a few, I have to say. So there's not. I think a lot of people see AI is like, well, it should all be like chat GPT where I sign up for a very low cost and it should be more useful and instead solve all of my issues. But really where it gets useful is like you kind of have to either build it yourself or have someone build it for you because it's just not to the point yet where you can tell AI to build a team of AI agents and it actually does it and they're actually useful. That's what I should make the next billion dollar idea. But it's, it's not there yet. So I think that it's like, you know, when, when there's expecting it to be a very simple answer to a very complex solution. And so like you said, when people envision, oh, the AI to do all the marketing, technically it does and can exist and it can be built, but really it's just like either you use off the shelf tools and you get the same marketing strategy as everyone else and it gives you very bland AI GPT answers and you think that it's so great because, oh, it made your whole marketing plan for you and did all that, but it doesn't really play out that well because there's no expertise in there and you didn't even know. If you don't know the process in the first place and want AI to do it, you're just going to get limited results. So I think that that's like, you know, thinking that they're going to go to someone who's just going to be like, I actually do have the agent that does everything for, you know, chief AI officer. I actually have it right in my back pocket and I can give it to you like that. Just give me, you know, 100 bucks a month and I'll just snap my fingers and it's installed and it's going to do all of that. Like, and I think it's kind of crazy because it's like if someone did tell you that, would you believe it? Like, I have an AI that will actually understand your workflows and do all of it and it doesn't need any customization or building. So I think the idea of like, you know when people like oh, so here's my exact workflow. I'm using Zoho and Trello and Asana and Outlook and all this. And then it's like. And they'll be like, oh, so I, here's what I want to do. And then we're like, yeah, can we, you know, that's something that can be built like here's the timeline, all that sort of stuff. And then like, oh, you don't already have it? I thought you guys made agents like we do, but it's like you have to customize it. It's like there's a lot of customization. So it's like. And would you really believe us if we said we actually have that exact agent right here? It's very cheap and it will be deployed in one day like that. It just doesn't make sense. But I think very non technical people think that like oh, if you make agents like you should already have the agent that I thought of before this call already built for me and ready to give to me. And it's just like that's not if even with AI, it does make things better, smarter, more dynamic. But it's, it's just not a snap of the fingers like oh, I'll just go boom, type into chat GPT make me this. And then there it is. It takes a lot of work. It a lot of times complex problems cannot be solved with no code tools either. So it's like it takes software developers, it takes mapping out your exact process. Like for example, so we have this client and they want a inbound lead agent. Seems simple enough. You can build similar inbound lead agents with no code tools even. But then when it comes down to is these very specifics of like, oh, their form for the inbound leads is actually a custom HTML form. Okay, do you need to have developers connect it? Oh, their, their channels that they communicate on are not really commonly used in the U.S. it's like telegram, Skype, you know, stuff like that. Where it's like. And it's like, oh, Skype doesn't really have a very friendly API if you go outside of Azure. So it's like all these little technicalities and then like okay, then how are you going to store the conversations with a custom thing. So when we create a new thread with a new client, we have to pair the thread ID to the contact in HubSpot. So you get down to all these little technicalities where it might seem very easy and like, oh, inbound lead agent, I'll just put it in this no code chatbot builder and boom, boom, boom. But when you're actually solving very custom and complex problems, it's like, it's, it's not really that easy.

So you've mentioned prompts a few times, and when most people, when people think about prompts, they think about going into Chat GPT and giving you a set of instructions. And are you saying that the agents, let's say at any level, really, are some of them running on what, like a prompt as most people understand it, or are they running on like a coding language for decision making? Or what's, what's, what's that scenario?

Yeah, so you, you still use, like, you still use prompts. If you're using AI, there has to be a prompt, even if you are doing it through code. So it's like, and that's another misconception, just to add quickly that people will be like, when you explain, like, agents are just code and stuff, people like, like their head explodes. And it's like all of its code. Like, if you use Zapier, make relevance, it doesn't matter. It's all code. It might have a cool UI for you to look at, but at the core of it, it's all code. And really code is the most efficient way for them to run. Not through a ui, it's just for people to see it so that they can envision it. So anyways, it it to answer your question, yes, they always need to have a prompt. However, the prompting is different for agents rather than like prompting Chat GPT. And a lot of people will think like, oh, I'm good at prompting ChatGPT, so I'd be good at prompting agents. But it's a lot different because with Chat GPT a, like, it learns from you, it remembers stuff about you. So you can give it a lot less context and it can come to a conclusion based on you using it for a long time. Another thing is that when you're prompting agents, not that it's like one shot necessarily, but it's like with Chat GPT, you could go back and forth five times and be like, well, I got a good output, I know how to prompt Chat GPT. But realistically, if it took you five times to do it, you're not as good as you think. Like, you know, it should be with agents. You need it to consistently and reliably execute a task in the right way every time that it's run. That's A good agent. And that takes a lot more in depth prompting and thinking of like a lot of edge cases really is what it is. So trial and error, thinking of edge cases, your experience of building other agents and knowing what the edge cases and where AI is likely to go wrong, and then writing that into a prompt that every time it's executed will do the right thing. So it needs to be broad enough. Like think of the prompt as your agent workflow. Remember we're talking about workflows. But this, it can be a more dynamic, but it still outlines the workflow of why it should make decisions when, what it should do, what it should use. Here's an example, et cetera. So it needs to be long enough that it gives it a full idea of like when to make decisions. But also another misconception that people have is they go to chat GPT and say, I want an agent to do this. And it writes out a prompt like this long. And it's just like so much fluff and so much like, just like random markdown bolds and stuff like that. And it's just like, you know, unnecessary. It's over explaining like, you know, the section for tone is like two paragraphs long and it's like, dude, you don't need to explain an agent's tone for two paragraphs. Like it's, you know, that doesn't make sense. So that's definitely a misconception. Like if I'm good at prompting chat gbt, I'll be good at that. But what it comes down to, yes, if you're using agents, even if you do it through code, you do need prompts to explain because that's how you communicate with AI. So that never goes away. But you can, you can use both. They can exist in the same flow.

So for anybody out there that's doing the back and forth and thinking, hey, I could, I could give the instructions for an agent. It's not as easy as, as you going back and forth asking it about, you know, what can I have for dinner? So I'm particularly interested in the acceleration of the space. Six months ago, or well, I guess maybe June of last year, I started getting really interested in the concept of agents. But not necessarily, I'm not a builder, I'm not anything like that. But the, because, you know, my, my perspective is, is that when GPT GPT 3.5 came out, people were afraid that AI was going to take their job, that 3.5 was going to take their job. And then once you get under the hood and you start using it a little bit, you realize, extremely helpful. But it needs me, right? Like it, it's not going to take my job, it's going to make, allow me to do my job better or somebody who knows AI plus my job replace me. That's more likely. So I think the people were expecting, like you've talked about, the one agent that does everything and that that's AI, which in theory is. And in practice we're seeing, you know, some examples for sure. But it's six months ago was a whole different world than it, than it is now and it continues to accelerate. What are you expecting to see happen in 2025? Because you'd look at any list about, you know, big breakthroughs in 2025 or AI's impact in 2025. Doesn't matter if it's Microsoft, Wall Street Journal, they're all talking about agents and you being, you know, neck deep in this stuff. You know what's possible and what isn't. But you also know what's being developed or what's being kind of hinted or promised that's coming. What do you realistically see agentic capabilities looking like? I don't know, by summer, by the end of the year?

Yeah, by, I guess exact timelines are hard to say. I would say that the future is like how every software is integrating their own AI tools. So that's how it started off of just like, you know, simple, like, oh, it writes an email for you in HubSpot and stuff like that. And like, you know, very simple. Like they're just, just using an LLM essentially. So. I see. I mean, we already know that like Agent Force is coming by Salesforce and HubSpot will have one. Like, I think that what will make them the most effective because you can't just guess everyone's workflow. And also a lot of these software tools don't really. They want, you keep everything in their tool, they don't want you going elsewhere and they don't really want it to interface. So there I think that every major software is going to start, you know, they have like, even in Gmail, it's like, oh, help me generate this email. But I think what they're going to start doing is then having agents within their own platform where it's like, if, you know, HubSpot builds an agent for HubSpot, it's likely going to be one of the best for that because it's going, it's, it's going to have access to all their API endpoints, be able to do whatever you just tell it to do. So I think there are a lot of these tools will go more to like chat with the tool. Like you won't go into Gmail, start a new email, put in the senders and stuff. You would just type or speak to it, just like, hey, can you message Chris and follow up on my email from last week? And then it's like, okay, boom, let's find Chris. Okay, boom, here's his email. Let's do it with the email from next week. So that's how I see it going in the, in the near future that yeah, you'll have good agents in every platform, but they'll only really like control that platform rather than do like an entire workflow. I think with the reasoning models and stuff like that. That'll be really interesting to see because that's kind of the main point of why agents are good because it's like they do some of that behind the scenes reasoning rather than just like question, answer. Like that's only somewhat useful. So reasoning models, just take that a step further. So I think that using those once you can use like let's say O3 mini to call tools and read files and images and stuff like that, they make it like multi modular or multimodal. It will be, it'll be extremely useful and it can handle more complex tasks. What those tasks are, it really, it depends on what you're trying to achieve. Coding ones will get better. We use some of those, if you've heard of like Devin. But yeah, we use that and it just, I mean the better the models get, the better the agents get. So it's like, you know, in the agentic framework itself it's like, yeah, teams of agents are good. I think those will become less relevant as the models get smarter and have longer context windows that they can accurately pull from and that they're better at reasoning. So I think that you could use fewer agents to accomplish the same thing. You could use the same amount of agents to accomplish more complex things. But yeah, actual like end goal of like oh, what is not possible today that will be possible then I'd say probably the biggest one from my perspective is just like sort of the just coding in general. Because like right now it's like we'll use stuff like Cursor and Devin and stuff like that. So like Devin will actually run in a virtual machine, go open up its own like let's call it like mini computer. And it does, you know, okay, read API documentation, write the code, but it's not even as good yet I would say as like a junior dev. So it's like. But as there's more reasoning, longer context, windows, et cetera, it will just be able to do more like one shot stuff where it's like you see stuff like Bolt and what's that other one like lovable or something like that. And so it's like make an app from a prompt and it's like technically you can and it's cool but really those are focused on the front end which is like most appealing to people because oh my gosh, I can see it. I made an app. Then you look at actual functionality, the back end, how it's architected and all that sort of stuff and it's just not, not really that like commercial. Yeah, yeah. And same thing with like Devin, you can't just be like, hey, make me this iPhone, tracker, habit tracker app, blah blah, blah, one shot, it, it's just not going to happen. So I think that there will be, be more stuff and like it'll go in baby steps like oh, next model Devin will be able to do a little more, A little more. A little more to the point where it's actually at like junior dev level where you say like, hey, here's, you know, you're, you're in my GitHub, here's our project manager. So you understand the goals of what we're doing. Here's like an SRS document of what we want to achieve. Ask any clarifying questions and then go do it and then it just like can actually complete that and you know, do do a full job rather than like parts of a job. So I think it will get there and you know, not that like OAI is going to replace everyone's jobs. I think that people that currently, I think it's going to make it more difficult for younger generations unless you're very good at AI to get a job. But I think that the people that are currently in the workforce that I don't think they're just going to get fired because most people, I mean, let's admit a lot of people even like CEO C levels and stuff like that don't know how to use AI at all. So they're probably not going to start firing their team because they don't know how to use AI, because they don't even know. So I think that it will, it will make people younger generations more hireable if you do have that skill 100%. And I think that there will be a large gap there because I think that rather than just Firing people's whole teams. They're going to start over the next five years slowly integrating AI, so they might just, rather than fire people, hire at a slower rate. But that can also in turn, if you have that skill and you go to a company, you're like, hey, I'm a data analyst also, I'm. I can build agents and I'm an AI expert as well. Then it's like, wow, double there. Because those higher up, older people, like, we don't even know how to do that. And they're like, yeah, I can get five times more more done than you using AI and I can show you how to do it too. Then it's like, wow, massive value. So I think that's where the kind of the gap is going to be. Not that everyone's going to lose their job, but it will be an essential skill to know like. Like using the Internet. Like, you wouldn't hire someone that didn't know how to use the Internet.

Sure. You know, so for those of you listening, like, pay attention to that. I think that's a really strong point. And that if you're good at what you do and you also have AI skills or certifications or training or whatever, it's not just about you coming in there and doing your job with AI, but if you can teach others on your team how to do it, you're lifting the whole boat. And that is very appealing to that. The hiring manager or whoever's making that decision. One of the things, you know, we, we hear a lot about with Salesforce and their agents and HubSpot and their agents, and I'm always skeptical. I come from a financial publishing background, so I know the shenanigans that go on in publicly traded companies are. Do you think that these companies, obviously, they know they need to do it, but do you think that right now that they're. They're saying it more for not necessarily capabilities, but more for hype about like, oh, we're agent forward or whatever.

Yeah, yeah, I think it's more for hype. I've definitely heard that, you know, although. Because agent force, like exists now, but it sounds like there is some trouble in getting people to adopt it essentially, and trouble rolling it out, which it's like, you know, is that gonna stop them? No, but I think of right now it's like a lot of talk and not a lot of, like, let's say case studies. Like, this isn't. This isn't actually something that all these companies are running right now. And like, oh, we're Seeing amazing results. It's more like conceptual and it's like to be expected too. Like you're getting into a new sector of a new technology and stuff like that. You're not going to immediately perfect it. So you know, it's just how it is. Just like, you know, AI like GPT 3.5 existed for quite a while before all these companies started to now add their AI tools into all of them. And since, same thing with agents, they're going to exist first and it's like six months to a year and then we're going to start seeing like, oh, okay, now all these have agents in them. So yeah, not, not like they're lying about it or they don't have it. I think that it's kind of just being worked on and more conceptual for a lot of companies rather than like actually in being used in practice and like deployed for customers. So.

Well, here's a question. If, let's say I was on Salesforce, knowing what you know, would you say, and I would imagine that there's a premium being charged for access to Agent force, would you recommend companies, I know, like you, you don't, you're not part of Salesforce's team or anything like that, but would you recommend companies be early in on this stuff to take advantage of even though it's going to be bumpy and there's going to be, you know, whatever misconfigurations exist today are going to be gone tomorrow. Would you recommend that companies go ahead and get in early so that they can start to change the culture or participate in the learning? Or would you say wait till it's stable?

Yeah, I, I would say just start, start when it comes out. Just like with anything with AI, it's like you're not really going to hurt yourself too much. Like don't tell the agent to delete your entire CRM. Which it probably would have parameters against doing that, but it's as long as you're not like if you don't over rely on it, think, oh cool, there's an agent for this. Now it's going to do all of my work. And then you're like, oh, why did everything get all screwed up? It's like, yeah, don't look at it that way, but kind of just test it out, see like what things is it good at? Because it might start where it's like it technically can operate all these parts of the CRM, but it's only really good at like pipeline management and sending emails. Okay, we'll use those get familiar with how it works. And like you said, AI develops quickly. So next week they might have a patch and they're like, this works now. So we were actually just talking about this today on another podcast, my CTO and myself, where it's like, like I was saying about Devin, like, it's, you know, a lot of people, especially developers, like to like crap on it and like, it's not good, it's not useful, blah, blah, blah, blah, blah. But it's like, you know what, when it comes down to it, it might save us like 10 hours a week on low level tasks. But however, it's like sometimes if a senior developer is working on a project, if you're saving 10 hours of their time every week, that's still like you're getting, well, your money's worth. We're getting like four or five times our money's worth from that at least. So it's like, you know, you could be like, oh, it can't even do that much coding. But it's like old things like writing simple JavaScript and API integrations. Like the senior devs don't really want to do that anyways. Like they, they, it's kind of this, like hate doing this. So. And then it frees up their time as well. But if you look at it. And then another thing, so another person on the podcast was saying like, well, if you look at it this way, like, even if it's not that useful now let's say it wasn't even that useful. Say it's like five hours a week just getting to know it and understand it and get better at using it, prompting it. And then when they actually make it better, they're like, hey, now it's using O3 mini and it's actually really, really good. We've been using it for this long, you understand how it works and sometimes it's like how you use it, you can be better or worse at using AI tools. So even if you're, the tool isn't perfect that you're using, if you get really good at using it and prompting it, like I was saying about the agent, prompting is different than GPT prompting. So things like that if you get really good at it with a tool that's not really that great, when it is that great, you're going to be amazing at it. So you already got the learning curve out of the way. Yeah. And there's a lot of stuff, even if they change the model on the inside, there's going to be a lot of stuff about it. That doesn't change and that remains the same. So if you're familiar with all of those and it just gets better, well, you're already familiar then with a really good system where everyone else that was like, oh, tried it and they're like, oh, it's not that good, it's not useful, I'm not going to use it. You know, I don't trust it, whatever, then there, then it's going to be actually good. All of their competitors are going to be doing amazing things with it and like, oh my gosh, I should learn this. And it's like, you should have learned it six months ago. Yeah.

So Jake, we're getting to the end of this, but I want, I want people to understand what they should look out for if they're interested in using agents, but they don't have the in house capability. That was our situation. That's why we reached out to you and your company to help us with architecting and developing and identifying opportunities for agents. What would you say would be some advice for somebody who is looking to engage with somebody like Agentic Brain, your company?

Yeah, I mean, I would say the first thing to think of is like we were talking about before, if you want something really complex done, the reality is just you're, we hear this all the time. I could do it with ChatGPT. Okay, why are you here? Why? You know, it's like the thing is it's not going to already exist if it's super custom. If it's solving a high level problem, it's probably going to take at least a couple months to make and it's not going to be $20 a month. Like, you know, you kind of have to weigh out like is your problem worth solving? And then just be wary of people that do tell you that, oh, I have the agent that does everything and I'll deploy it this fast and only for a very cheap price. I would be very skeptical of that because it's like, how would you already, how would that person already have exactly what you want? I think that it's going to be a long time until agents can build other agents to solve very custom and complex tasks. So that's, that's kind of what I would say is kind of consider like, and if your problem is super custom and not that high value, maybe it's just not worth solving. Like if it saves you $1,000 per month, but is extremely complex and custom and stuff like that. Like, to be honest, there's some things that just aren't worth solving. With AI, however, if you have something that is really complex, let's say for example, a client we just worked with, they're like one of the things that takes 25% of their time because they're doing like insurance for like, like condos. Essentially just data entry, moving it back and forth takes like a quarter of their time. So again, is it kind of custom and like, I wouldn't say insanely complex, but a little bit custom. A little bit more complex than you'd probably do with no code tools? Yes. However, to get back 25% of your time is massive value, so it's well worth it. So kind of look for that ratio of like, if it's, you know, determine how custom and complex it is and then look for things that are going to provide a huge value to you. Then go look at getting custom solutions made for them. If it's very simple, not a high value, maybe you know, just find a tool that can already do it. That would sort of be my advice. But if you're, yeah, looking for, looking for big complex processes handled, that's when you would want to actually have custom agents for yourself.

So do you guys have bandwidth?

We do have bandwidth, yeah. We're growing our team.

Okay, awesome. And for everybody listening, full transparency. I engaged with Agentic Brain to do this same process for our own business. We just got through doing kind of a proposed roadmap and everything. And the suggestions are very exciting as far as the impact that it would have on our business. So if somebody is interested in exploring this for their own business, what's the process for them to connect with you guys and do go basically. Basically go through what I'm going through.

Yeah. So you can either go to our site, agenticbrain.com or you can always contact me on LinkedIn if you want to drop my LinkedIn URL as well. Somewhere connected to this as well, that's the best way to get into contact. There's a calendar on our site as well. And you know, what we really do on the first call is help people determine. Like, because to be honest, we don't want to work with everyone and that kind of ratio that we're looking at. If your job is very expensive and not high value to you, a lot of times we, you know, we can, we help determine that in the first call. Just like you might have better luck using these tools instead. Custom agents might not be for you. But then also there are some when it's just like, oh, wow, this would happen such a huge impact on your business, it's definitely worth moving forward. So that's, you know, that's kind of our process of what we help people determine on the very first call with us.

So for those listening, like, I'm not working with a whole lot of agentic companies, but I don't know that we need to. The experience with Agentic Brain has been what I was expecting and what I was hoping for. So endorsement for reaching out to Jake and his team, especially if they have bandwidth. I think that if your experience is like mine, this won't be a mystery anymore. You'll have a very clear path forward. So, Jake, thank you. So obviously, you know, I appreciate it sounds like you're a busy podcast day, but I appreciate you taking the time to share with the audience here using AI at Work. And we're going to include obviously the contact information for the site. I want people to follow you on LinkedIn and see the type of stuff that you're doing. And for those of you that are in our community, I'm going to be talking over the next few weeks about some of the suggestions that they made and just to kind of give you guys an idea of what a business like ours, how it can benefit from, you know, agents in the hands of experts like Jake and his team. So, Jake, thanks again, buddy.

Thanks for having me on.

We'll see you guys on the next episode. 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 you get started using AI at work. That's www.chiefai officer.com. thanks to our producer Evan Desalnier for making this episode possible. Follow us on Twitter at the handle usingaiatwork and visit www.usingaiatwork.com for free resources to help you harness AI in your role.