Agentic AI For Marketers With Sebastian Chedal

We map the shift from chat-based AI to real agentic systems that can run workflows, manage memory, and publish results without constant human babysitting. We break down how to design AI agents like jobs so they stay grounded, cost-controlled, and useful for marketing, SEO, and software work. 

• tiering AI use from assistance to automation to full agents with memory and tools 
• writing a clear job description with input, process, and output formats 
• building an SEO research agent that generates briefs and scores novelty 
• grounding agents with state, scripts, and tight context management to cut errors and token waste 
• why models ignore guardrails and how system design fixes repeat failures 
• plan mode tradeoffs and where to focus review time: planning, code, or tests 
• orchestration patterns like dispatchers, specialist sub-agents, and clean bridges between systems 
• security and scalability realities for enterprise-ready software built with agents 

Guest Contact Information: 

linkedin.com/sebastianchedal

chedal.org

More from EWR and Matthew:

Leave us a review wherever you listen: Spotify, Apple Podcasts, or Amazon Podcast

Free SEO Consultation: www.ewrdigital.com/discovery-call

With over 5 million downloads, The Best SEO Podcast has been the go-to show for digital marketers, business owners, and entrepreneurs wanting real-world strategies to grow online. 

Now, host Matthew Bertram — creator of the LLM Visibility Stack™, and Lead Strategist at EWR Digital — takes the conversation beyond traditional SEO into the AI era of discoverability. 

Each week, Matthew dives into the tactics, frameworks, and insights that matter most in a world where search engines, large language models, and answer engines are reshaping how people find, trust, and choose businesses. From SEO and AI-driven marketing to executive-level growth strategy, you’ll hear expert interviews, deep-dive discussions, and actionable strategies to help you stay ahead of the curve. 

Find more episodes here: 

youtube.com/@BestSEOPodcast

bestseopodcast.com

bestseopodcast.buzzsprout.com

Follow us on:

Facebook: @bestseopodcast
Instagram: @thebestseopodcast
Tiktok: @bestseopodcast
LinkedIn: @bestseopodcast

Connect With Matthew Bertram: 

Website: www.matthewbertram.com

Instagram: @matt_bertram_live

LinkedIn: @mattbertramlive

Powered by: ewrdigital.com

Transcript

Welcome And Why Agents Matter

SPEAKER_00 0:02

This is the unknown secret of internet marketing. You're inside a guide to the strategy of top marketing union to craft the competition, ready to unlock your business full potential. Let's get started.

SPEAKER_03 0:18

Howdy, welcome to another fun-filled episode of the unknown secrets of internet marketing. I'm your host, Matt Bertram. Uh, I would tell you, my computer might be uh you know maxing out right now. I got uh too many things running. We got some uh clawed code going, we got uh a couple different agents set up, and so I probably need to close some browser windows and turn some stuff off. So uh if there's a little lag, guys, apologies. Let me try to shut some of this stuff down. But as the conversation has moved more and more towards the agentic realm, and we're talking about LLM visibility um and just kind of how to think about all these things. I thought I would bring somebody on that is deep in uh AI agents, uh setups, audits, uh setting up that AI agent to help you uh help with your workflow. Uh Sebastian, welcome to the show. Thank you so much.

SPEAKER_01 1:13

Pleasure to be here.

SPEAKER_03 1:14

And Sebastian's with uh fountaincity.tech. Uh you can go check it out and uh I'll let you talk a little bit more about it at the end. But you're doing everything AI, Sebastian. So I thought I'd bring you on.

SPEAKER_01 1:26

Yeah, thank you so much. There's a lot to talk about here. So where would you like to start?

SPEAKER_03 1:32

Well, yeah, the I mean the market's moving really, really quickly. I I felt like open claw was uh a little bit of a distraction. Um we were setting it up, but it was having some different kinds of problems, and we've just pivoted back to quad code. Um and uh just kind of now they leaked the harnesses. So like I think everybody's gonna have a pretty good harness. I know um, you know, what Facebook's doing is really interesting. Like we've been using Mantis a little bit because you got the you got the rag recall, quote unquote, from all the Facebook ads. Uh, so that's helpful. Uh, but it would be good to maybe talk about how you like a mental map on how you look at all these different like frontier models and how you're maybe using them together. And like even like uh perplexity, I've been playing around with uh Comet and Computer, and it's really agnostic, right? So it'll just use whatever's best to kind of build it. It's like a yeah, Po 2.0. Uh, if we're moving from uh the chat bot side to really the the agent side of things, I would love to just kind of how would you paint the picture to someone that is around this, but maybe is not as deep in it as you are.

SPEAKER_01 2:47

Okay, yes. I think what's coming into my head as you're talking about all those different things is first I would classify the different usage into tiering as kind of a way to kind of focus then in later on certain aspects of it. So AI assistance is kind of tier one where you're just using AI to help you along with whatever you're doing. And sometimes that might even look at how people are using cloud code, but certainly perplexity or a comet browser, you know, you're and then the next level up, uh, tier two, you're using AI to perform certain automations that could be on a clock, that could be triggered. And the workflow automation itself might only have AI as one of those steps in the process, but there's AI, you know, somewhere in there, otherwise it's just a normal workflow automation. Tier three is where I start thinking about the AI agentically, meaning it has memory tools and agency, meaning it can take its own actions either on a schedule or triggered as well. And then if you want to go beyond that, there is levels above that, but already at that third level, you have within that two different categories already as well, A and B. A would be it's driving my computer, right? Which is more like co-work or perplexity. I forget the name of it, when it's driving your computer computes, I think is that what it's called from uh perplexity computer.

SPEAKER_03 4:07

Um, and then like claud co-work is the what's driving uh is is the shared kind of uh online component. And then uh claude uh code is what at least how I have it set up, is driving mainly my computer. And so we're starting to put things in um like uh Apple Minis where we're we're playing around with like ChemKey 2.5, uh and and we're we're because the token usage, unless you really manage it, like putting like a slash context or something in front of it, it will just burn through the tokens so fast. So, like tokage optimization has been the biggest area that we've been trying to tinker with because it, you know, you get it to do a lot of actions that get expensive.

SPEAKER_01 4:53

Oh, totally, yeah. And so that that second category is more the always-on category that could be, like you said, like a Mac mini in your house or something like that. We, I mean, we're we're building these systems for clients, so that's not really an option. So we put them up in the cloud, so it's like AWS EC2 servers or stuff like that, you know, or Google, you know, whatever flavor of system the client is in, we'll just embed it into that ecosystem. Uh within, but yeah, and well, you just talked about cost management. I think there's different so talking about open claw or these kind of systems that have agentic capabilities into it. There's so many different ways to build that system. I think the first way people think of it is, you know, maybe they're experimenting with it, they do conversations with the system, give it some abstract goal. The AI goes off and tries to accomplish that abstract goal, burns, you know,$50 a day in your wallet of token usage, and you're like, oh my god, I don't even know if I'm getting value out of this, you know. So we actually do so we've been building a lot of these autonomous workers or autonomous agents. And there's uh we like to think of the agents kind of categorized also in terms of those different tierings that I just mentioned. So we might have one agent that really has a very specific workflow.

SPEAKER_03 6:13

Are you what are you using for orchestration? Are you using like Langchain or something like that? Or like I'm trying to understand what the tech stack is because I'm starting to use quad code and hooks and skills and like um it's spinning up other agents underneath, like whatever it's trying to do. I I think I'm probably teetering between like uh if you framed it up, between two and three, because we're just setting um triggers like daily or like autonomous, like okay, go go run the script. And then there might be a agentic layer on decision making at some point. But like, can you back up and break down like what is the tech stack that you're building on? And like what are you setting up, or maybe even give an example of something you're trying to solve for a client?

Designing The Job Input Output

SPEAKER_01 6:58

For sure. Yeah, I will add one pre-step to what I'm about. I'm gonna answer your question. I'm gonna say one thing first, which is I think the most important first step is to figure out what you're trying to do and describe that. I like to think of the coming up with a job description of what what is gonna be built. And then within that, really think about the input. What is the context? What's gonna be the input for each job? What is the overall general input, like you know, first day training on the job, what you need to know. Then what is the process? How does the sausage get made in the factory? That's a really important step. That's where your quality comes in and not just volume of slop. And then what's the output format, right? And so once you have those three, then you can start thinking about the architecture and how you're going to accomplish that goal and what tier level you need and combination and so forth. So, with that in mind, let's say that. So uh so we have one agent, for example, that does SCOGO research and builds content that feeds into a content pipeline with ideas and so forth. And so within that job description, there is a whole need to understand the context of the business at heart, but then also who are the competitors, uh, what platforms do we want it to research every week. So we have it researching all posts from all competitors, plus Reddit, plus X, plus Substack, probably a few, and then just a general uh mentions search across the whole web of who's talking about what. All of that's compiled along with you know documents that describe the objective, and then all the Google Analytics is read, the Google Search Console, keyword performance, AI ranking performance is also tracked, like all of that's input, and then from that, briefs are generated according to a format of how we want the brief to look. Those have actually have strategies as well of what kinds of briefs, and then that's the ultimate output at the end of that agent. Is like, okay, here's the things that should be written. And we do also score them in terms of um I forget the name of it, but it's a scoring that we use to determine the tier one is that it's very novel, unique content, right? So it's it's it's authoritative. Tier two is the uh the research is going to be novel, so no one's researched this before. Tier three is the framing is novel. So the research and the content isn't, but the way we're framing it is, and then tier four is what we call commodity content, and then depending on the client, that type of content may never get go further, right? That's like a barrier, like because that client doesn't want commodity content. Commodity meaning if it disappeared from the web, no one would ever blink an eye, right? That has no intrinsic.

SPEAKER_03 9:34

And so you're you're using you're using like uh APIs, MPCs. There's kind of some new stuff that's coming out, as well as maybe some like uh uh you know, a data lake or even like a rudimentary Google Drive.

SPEAKER_01 9:50

And so yeah, I can go to the technical stack next, but I think it's important to frame like what's the job first, because from that, then you can start moving more and more into like okay, the nitty-gritty. So in order to achieve that, there has to be a memory management system. So that could be SQL database or obsidian QMD. Uh really it doesn't have to be RAG necessarily, depending on the volume, could be just relational. And so there's that piece of it. We're building it on an EC2, like I said. You need to have a clock that generates the reoccurrence. That could be a prompt job, could be open claw, you know, could I mean we're um uh so from from that clock then that's creating the occurrence, you've got the memory that, then you have the skills that are developed in order to create consistency of behavior. And then we uh there's a blog post on my website that talks about four things to really look at when you're building, which is if I remember them all off the front of my head, uh ground it. So you want everything to be state-driven. You don't want the AI model to just kind of remember what the state of something is and track it in a markdown file, right? You want to put it either in a JSON or table or something for state-driven decisions. You want to script it as much as possible. So you really don't want the AI to be thinking through how to do things over and over again. Even with a skill, I it's still not um, you're still gonna create error rates because AIs are probability machines, right? So you want to create scripts that the AI actually triggers. So a skill triggers a script, the script is a set amount of code, the code runs, it's not AI anymore, and it's also cheaper, runs an operation, stores states into your data store, and the AI is just taking data back and then doing the only the step, the only step that AI should be doing is the interpretation. Like, how do I interpret this information? How do I make a qualitative decision in this case based on this data with this context? You do have to manage also context because like the more skills you have, the less input tokens you have. So that's like an interesting balance right now in the ecosystem. Um but then so you so that's uh that's the memory agency, and then the skills and tools. That's really the main components of the system, right? And then from there, everything else is just driven by the input output layering. So for us, those briefs are an input to another agent, which then takes it further. And the ultimate final output is that it just publishes it on whatever website or social media platform is connected to the system or newsletter, whatever. So that's the ultimate output. Uh, and then the input layer that can be that research I talked about, but it could also be transcripts or video recordings, or it could be even other AI agents. So we've coupled also AI fully agentic systems that build software, and then those agents not only build the software but write case studies about what they did, the problems they found, how they solved it, and so forth. All of those lessons learned get fed into the system that create novel content that is then tier one for novel content, right? Because like it's act it's not just research anymore. It's not, and so you've got that you can actually have highly novel content going through your system end-to-end, which is like you don't even need a person anymore to do some of that novelty in the content, especially if it's like you know, research driven or things like that.

The Future Compression Of Work

SPEAKER_03 13:09

Yeah, that that is actually something that we were talking about earlier today. Uh, everybody listening, it's Monday. And so we were we were talking about how to uh pull some of that in. Uh and we haven't quite figured that piece out yet. But that would be great to uh talk to you more about that, because that's actually what we're looking at is like, okay, the data and the data points we're producing, how how do we turn that into case studies? Because that's um unique data, right? Um, it's it's data that we have access to. And, you know, so that's very, very cool. Very cool. Well, that's I think that that's what you talked about is where the market's going. And I think that uh a lot of agency owners are trying to kind of figure out that and are at different places and getting that done. Where do you see, like if you tried to project out, because I know it's hard because things keep changing, um, where do you think everything's ultimately going at at the end of this? I I've been listening to some different people that's in time or in 10 years, like what what time? Yeah, well, uh I you know, everything's moving pretty quickly. Like, let's say like three years, like where do you see everything converging in in three years? And like, where's that suppression compression happening in your mind? And what is the world gonna look like?

SPEAKER_01 14:27

So the compression really is between idea and delivery. And another way to look at it would be problem solution or opportunity and action. Whatever you want to frame, you know, the beginning and the end point, everything in between is compressing into a slice that's gonna get smaller and thinner and thinner. There's certain areas right now also that people think are relatively protected, domain, you know, domains that humans have an advantage over AI. And uh, I I think though a lot of those are just rugs that we're standing on that are gonna get pulled out from underneath us over and like subsequently one after the other. It's a very similar trend that we've seen with computers. You know, if you look at, you know, first we thought it could never beat a human in chess. That happened with what was his name, Kasparov. Then we thought it would never beat Go because the domain spaces were too big, and then that happened a few years later. Then we thought the Turing test was unbeatable, that was beaten in 24, right? So it's like we keep putting lines there and we think, oh, it'll never pass that. I think the really true so I think humans really have four main areas that are kind of the main areas to be at play with the with the direction things are moving, and these the four areas really to be thinking about for how humans are involved. The first is direction. That could be direction of the whole business, but it could also be direction within your domain, right? Within marketing or within design. Some people call it taste, but you know, it's it's ultimately like what is the difference between option A and B? Like, I guess even more abstractly, you know, if my kid came to me and said, Should I be an artist or a doctor? And it's like, well, what's important to you? You know, like these are so abstract choices that someone ultimately still needs to make uh call it a taste decision or direction within that thing, yeah.

SPEAKER_03 16:11

So taste, trust, governance, um, and then there's like one or two others that uh I've started.

SPEAKER_01 16:19

Yeah, trust I would connect to the second quality, which is the interactions layer. Okay, people want to interact with people. And so whether that's service or sales or support or teaching or understanding or empathy and trust and all these kind of things are kind of on the interaction layer. And so that's not going away, you know, that insofar that people are actually good at that, right? Certain types of interaction layers will eventually not like people won't even want it, you know. The time will come where you go on and you do a phone call for support, and if you get a person, you're gonna think, Oh god, not a person. I wish it wasn't the guy I was talking to, because it's gonna be so much better. Like, why is this company still have a person doing this job, you know? So, because you're you know, for things that are more like transactional or finding information or getting an answer, right? Those kind of things again are gonna just compress to the point. And I've already experienced that, by the way. Like there are certain companies out there that have figured this out. And when you do support, like I've done tech support with this one company, it's all AI, and it's at first I was like, okay, how's this gonna be? And it turned out to be amazing, like way better than like super fast answers, really accurate, and knew every single question I asked, like instantly, and I was on and off on that support call, like within minutes. Normally that'd be like a 15-minute call or more. So me ask my supervisor, right? Like, is the things that like would don't need to happen anymore. Anyway, and then so those two are kind of horizontal type jobs. And what I mean by that is they cut across disciplines, so there's compression that's on that compression that happens between idea and delivery will be most experienced by people in those two horizontal domains. So right now, you know, if you want to build a website or something like that, it's like things are very staged between people. It's like, oh, you have an idea, okay. Well, let's get the designer involved to design it. Okay, let's get the UX UI study. Okay, let's let's get the programmer involved, figure out how long it's gonna take. Okay, let's do who's the tester needs to come up and come up with the test. Like all of that's compressing so that the people in these two domains of direction and interaction get a much quicker experience from beginning to end. And it's compressing already as we speak. The other direction, the other two uh human involvements are what I call vertical instead of horizontal. And they're just they're the builders and the supporters, so they're the ones building and supporting and maintaining this infrastructure that is the thing that is the special sauce, right? Like it could be your differentiation, could be uh, you know, all of your moat that's created from your data, your specialized processes, the proof that you've run this a hundred times and on this particular type of problem, and you've gotten these outcomes, which are much better than anyone else, you know, whatever that is, they're the ones building that secret sauce, that value in there. And you know, that's all the like the building of the agents and the optimizing of them and the memory management systems and helping the AIs to replicate and to fix their problems, and they're the ones behind the glass looking over the machine floor at the robots that are building the cars, right? Like when we went from because we're kind of transitioning in that in the agency world from the world where you've got 10 people on the floor who are each putting together a different part in the car, and right, and some people are trying to figure out how they can get a robot to pass them a tool on the floor, right? That's like AI assistance. Can you give me the right tool in anticipation of my need? And then the shift with the agents with agentifying the whole thing is like, no, no, no, just get off the floor. We're gonna put you behind glass, and you're gonna look over the hundred robots. Five people look after the hundred who are building the cars. And there's still work to be done, there's a lot of work to be done for those people, but it's a very different kind of job, is what you know we're seeing. It's architecting, uh, you know, translating things that are in people's heads to now be quantifiable into a process, right? So it's a lot of process engineering.

SPEAKER_03 20:10

It's it's agent, well, it's human on the loop, not in the loop anymore.

SPEAKER_01 20:15

They're just yeah, or human behind glass is what it's like this year, right? It's like it's not even on the loop anymore, it's just behind the glass and looking at the data.

SPEAKER_03 20:23

Yeah, well, it's just like auto-approve, like set set the guardrails and like auto-approve.

Why Agents Ignore Guardrails

SPEAKER_01 20:29

It's we have yeah, we use tiering for risks, and that's how we determine auto-approval. So things that are low risk, and that bar keeps going up. What is low risk? Yeah, those are auto-approved, medium, it finishes all the way to the end, and then there's a checkpoint at the end, and then high risk, it's a gate at the beginning and the end, right? So that's how we kind of there's different ways you can measure it, but that's how we do it.

SPEAKER_03 20:50

And and I, you know, tell you tell me you're like I I feel like I have a I have an agent set up that's just watching, like making sure that even though here's the guardrails, like sometimes they do crazy stuff. Like, why why is that? Why why would you say that if you give it definitive like guardrails, it sometimes doesn't fall it. And I have to have like an observer uh agent watching that agent, making sure that it it it doesn't, you know, hard code in um you know uh API key or so, you know what I mean? Like, why does it do it even if you tell it not to do it? Like like I I would I would say it's like data hygiene, but you know, if this is a new project, like there it's like it shouldn't be an issue, but sometimes it is.

SPEAKER_01 21:37

There are different reasons that could be happening. I'm not gonna be able to tell you one reason without actually diagnosing your particular case. But one reason, I'll just go through a bullet list and you can dive into any one of them if you want. So one of them could just be the model mismatch. You know, you might have a model that's just not good enough for that. Like if you're using haiku or something equivalent to that, they're very pattern based. They're looking for patterns and they replicate them. Rather than thinking through problems. And then the mid-tier models like Sonnet, they're better, but they can definitely, I've seen them definite. What you're talking about, like hard coding values in, like, oh, your test case says that in this test it should come up with this answer. So I'll just hard code that value with a magic number in the code, and all of a sudden your pat your tests pass. And it's like, oh my God, I can't believe you just did that. You know, that's not the goal of the testing, whatever. But uh another category, so you know, you might just need a stronger model, right? That sometimes that fixes things. Uh like Opus is very good at following instructions, and uh, we've been experimenting a lot also with GLM5, uh 5.1.

SPEAKER_03 22:39

So you just think if they're getting get better, like they're just gonna like that's one, that's one vector, right?

SPEAKER_01 22:44

Of like possible reasons. Another is context and memory management, and and you could also include skill management and tooling as like all part of that category. So if you see, for example, and this, oh, and then the the next thing which ties into this is how do you approach problems when you see them? So whenever we see a problem, we always ask ourselves, you know, the the question isn't why is this model making this problem? It's what can I do to improve the system to make sure that this problem doesn't happen again, right? And so that framing is really important to take. And sometimes I see people just get start writing angry messages to the AI or tell it more sternly how it should be doing its job, you know, and like that might get you some results, but it doesn't mean it's it's gonna happen again to you, right? Because what you're not doing, you have to really be thinking about system design. So, you know, so take a crash course on like how to be a good system designer and look at what are you know what are the top five, 10 things you need to learn to be that to embody those qualities, and then that's how you should be approaching the problems. So if you have an AI that's not following the instructions consistently, maybe you need, you know, you can even ask the AI, do you have a good skill for this? What skill are you using? Or you can look at the log of how it did it. Maybe it tried something and it got an error, and then it tried something else, got an error, and then it tried a third thing, and then it found the answer, and then it took that, and then you can kind of watch what it did. And if you're seeing that it's kind of bumping along the way like a pinball machine, you can go through that with the AI afterwards and you can streamline it with a skill that tells it immediately the correct way to do it, so it's not having to bump around like a pinball machine. Sometimes people, the prompt or the system prompt or the thing that they're writing is not actually phrased in the way that is the most effective. As models get stronger too, you really won't be focusing more on why you want to be doing something, which is often forgotten, that you need to communicate that to the AI and what the desired end goal of it is, and not just the steps, like I want you to do step one, step two, step three, step four. That can work, but especially as you get to stronger models like Mythos that's coming out, you want to really be making sure that you're guiding these AI systems with goals and purpose of what they're trying to achieve, because that helps the AI a lot to make sure that it's not going to shortcut you with like a magic number in there. Because if you're telling it like we want to make sure that this works for all possible values of this variable, in this, you know, here's going to be some test cases, make sure it passes those tests. That's very different than saying, here's a test, uh the answer should be 1.5, go test it. And then the AI is like, oh, we want it to be 1.5, let's hard code that thing in.

SPEAKER_03 25:28

So so I saw a podcast interview with the founder, and he was just like, I use plan mode like for almost everything, right? To to to to give it those objectives to answer those questions, which I think are extremely helpful. One of the things you made me think of was like, how are you solving like persistent memory? And you you mentioned obsidian earlier, and and we've started to toy around with that. I'm not like an advanced user, um, but I I would love to know kind of like how you're solving the persistent memory and kind of like connecting it all together because you only like you said, you only have so much in the context window, and when it compresses it, it kind of it it loses stuff, and you got like the heartbeat component. Like there, there's just like uh with the so many things I can say now because you keep mentioning new things. I know, I know, I'm sorry, like clean it up for me, clean it up for the listeners of like I do.

SPEAKER_01 26:26

I want to also just touch base on plan mode because plan mode is awesome when you want to be building things, but there's a very big difference between planning, which is what to build something, versus using something. You can't use plan mode to use an identification. Yeah, so just want to make sure that that's clear. But in terms of building, I there's a lot I could say around plan mode, how to make that even more effective. Let's pull it out. I don't know. We just want to we don't want to.

Plan Mode Versus Review Strategy

SPEAKER_03 26:53

No, I don't know what's let's go down some rabbit holes of like and let's like actually give some context. So let's talk about plan mode and we'll move into the okay.

SPEAKER_01 27:02

So there's first of all, there is a big debate right now between where to get the most leverage and plan in in the planning and execution when you're doing agentic coding. So you're using a team of agents to write and execute the code and self-test and all that kind of stuff, right? So it's not AI assistance where I'm just having it write one function at a time, one class. It's like doing the whole refactor or the whole project. Um, those three points of contention are should humans be spending most of their time during the planning? Should they be spending most of their time during the code review? Like some people are saying right now. Some people are changing their mind too. They some people used to say, read all the code. Oh, that's right, sorry, read all the plan, don't read the code. And then those same people are now saying we made a mistake. You don't don't read all the plan, read all the code. But there's a third place to also, there's a third philosophy or a third camp, which is review all the tests, right? So that's where you really need to be putting all your energies at the end point, is the testing. So you're kind of doing a test-driven setup. So the I do think personally that that contention of where to put the effort between the three, I think of it almost, you know how you have those min-max absolute points on curves, where if you had a two-dimensional graph and there's like a bell curve and you want to kind of be in the optimal value with a peak somewhere around there. But because it's three points, like time, uh sorry, testing, code, and planning, it's actually in like a three-dimensional space between those three, and like where do you put most of your effort? But I do think that that whole question will go away as models get bigger, better and bigger. Um, there's still so much to be done around memory and context, which you're hinting on. I think that's really a big gap right now that companies will most likely start having to focus on a lot more in the coming year or two. Um, but in terms of the planning, when you're doing it, so the way we build plans is first we write up a uh either a product design requirement document or an ADR, an architecture design requirement document. Depending on the refactor or the code that we're writing, that could be somewhere around two to six thousand lines of planning that is created. That's a whole process. This is where you're putting a lot of time in, at least for us, is the planning. Our planning time is gone, it's like way higher than it used to be, right? We used to just like write a write some specs, figure out some stuff, and start coding, right? Now we're spending about double the time than we used to historically on all of this planning stuff. And so you so we do that first planning up front, then we take that document. We're not done with planning, because then we take each phase or each section of it, and each of those is then broken down. Uh, one of the things we tell the AI is write it as if a junior was gonna write this code, right? Like how how give the junior enough information that they know exactly what they need to do for each of these steps. There's more to it too, because you really have to think about edge cases and how you're gonna test things and validate, and how are you gonna get the AI to do all of this work without a human needing to be involved with these steps? Sometimes you have to build a bridge. There's like mini projects that have to happen to just build those bridges or those self-validation loops. So you build all of that out, and then from each of those documents, then you build uh work order collections, which are breakdowns of uh context that is a living document, plus a planning, which is a living document, and then actual individual tasks, which is also living documents too. Living, meaning that they're edited as they're being used, they're not just static. And then those collections, and there are different frameworks, but this is kind of the most simple one. Those are then passed on finally to a team of agents that then has their own um state-driven control system that I was talking about to make sure that things are grounded so that and they have antagonizing objectives, is also important too. Like you want the tester whose goal is to find as many bugs as possible, and they get pats on the back or best fit points, or however you want to look at it. And then the developer wants, of course, to code successfully, and then the task manager wants to make sure that everyone's proving that they're doing their job so they can mark them all off. So every single role in it has their own objectives, and some are antagonizing. And then they go through all of those, self-check and report, and then you just need to, and then from there, you as you sophistic make the system more and more sophisticated and resilient, you can accomplish larger chunks of work at a time before a human needs to even go and start reviewing it. So it might run for five minutes in the background when you're building confidence, but then eventually it's running for like 30 minutes or more in the background as it just crunches through these different documents in series. Um now, planning one-off planning mode in claw code, that I would say I kind of view it as like it one step above AI assistance, because what you're doing is you're basically doing a one-shot job in in air quotes on a particular function or feature or mini refactor. But there's a limit to the scope and scale that that can accomplish on its own. But it is definitely way better than just a one-shot prompt in it. But even the plan mode needs supporting memory and skills and things like that in order to make sure that the plan is grounded in truth and doesn't just have assumptions, surface-level assumptions in it. That's that's kind of the devil a lot in the detail with a lot of the plan mode, is that AI has is sitting on top of a very large code base, reads three files, but doesn't read the other 10 it needed to read, does a plan based on assumptions and what the other files does, and all of a sudden it's completely gonna just screw you over when it goes through. You know, we've seen seen those problems happen. So, you know, plan mode can be amazing, but it can also be insufficient for what you're doing.

SPEAKER_03 33:00

No, oh yeah. No, I I I think it was kind of like a bare minimum. Like if you're gonna like just a one-shot prompt, like if you if it's not really elegantly uh engineered, uh it's not gonna give you the output you need. Because there's a lot of conversations, I guess, on social media of like all these tools that are just kind of wrappers anyway, and it's like build me a whole app, and then it like breaks. And it's like you can't build a whole app with one line of code that and then and then you got to think about well, how many users are gonna go through this app at the same time? Because you know, if you if you scale it up, you got one one function processing at a time, like it's not gonna be super useful. And so that actually, you know, I know I'm pivoting all around. I have so many questions for you, but that's been some of the debate with a couple client projects that I got pulled in on. Is okay, like we want to vibe code something, um, and it's functional, but there's some security concerns, what kind of data is gonna go through it, how many users it's gonna be, and then like, okay, can like base 44 or so some of these tools um uh replaced an enterprise system, right? And it's like, no, use it to create the MVP and then at a certain point switch it over to a developer. And I'm starting to see with with these agent systems, it's like, well, they can code it up, like, you know, like like okay, maybe you're not using this tool, but these agents, if if, if they have sufficient information, and I get it goes back to the the memory. I feel like the memory is that that big issue right now, but I mean they can code it in whatever you want to code it in. Like you can build a plugin, you could, you know, build it with a similar back end. Like if you're building a WordPress site, you could build a similar looking WordPress backend, or hey, let's just use um the the the command line or the terminal to do all the work and let's not ever touch the site. Like, like there's a lot of these discussions of like, I feel like it's when um like digital cameras switched over to like a new technology, and it's like like you you gotta make like a hard switch at some point to create different processes and do things differently because a lot of people are augmenting the old way, but there's like a new way of doing things. And like, where is that switch over point and also the maturity level of of what these things are doing? I, you know, like do you have to hand it off to a full stack developer to build it out? Or, you know, can these agents build it where it's enterprise ready? Like, what's happening with SaaS? I there's a lot of questions in there, so feel like to answer whatever one you want.

SPEAKER_01 35:50

Sure. So what I would say is, I mean, talking about that last thing you were just mentioning around the full stack developer. Yeah, the way we've seen so we made a rule for ourselves first experimentally back in December, which was can we do this new project never writing a line of code? And it was a pretty it was a hydraulic system simulation code refactor, you know, a meaty thing, not just like something you could just spin out as a five code thing in it in an afternoon. And back then it was also pre-Opus 4.6. I think it was 4.5 at the time. There's something magical that happened with 4.6, but that's comes later in my little story. But man, it was like the first month was like painful because we were just we had this rule and we're trying to go through it, you know, regardless of how to learn how can we get this thing to to be as good as us doing it by hand, but not right. And so, but now with the way things have progressed and and all the things that we've learned, and then also 4.6, which did like this magical kind of frosting on the top and like really helped brought it even more to reality. I at least for us, like it's kind of over the whole idea that you write code by hand, you know. Of course, if the client comes over and it's just like, I need to change something quickly, I mean, maybe you just change the line of code, or maybe you just one-shot that one little thing. But any kind of major refactor or major project, it's we're seeing that it's done way faster and has less problems than when people are doing it. By less problems, I mean less bugs, you know. So, I mean, that's kind of a big kind of milestone. Is that and then by and less bugs can compare to like let's say the average programmer, right? I mean, of course, there's gonna be senior coders out there who are gonna still beat these AI systems today. I don't know how long that's gonna last, honestly, like probably a year from now. That might not even be true anymore. But at least right now, you know, certain senior devs can probably outcompete the AI still, but probably not on speed, but more on you know, bugs and broader thinking of impacts and things like that. So there's there's still some a little bit of a moat there. But um, but yeah, in terms of enterprise ready software and so forth, you do need somebody who is that brain who really is thinking through, you know, what are the security levels? How are we gonna test this? How is it gonna integrate? When we're done with feature one, what's coming next? How do we not paint ourselves into a corner? All this kind of like broader level thinking, the AI is not capable of doing that. AI is terrible at asking tangential questions. You know, it's like the AI will take a task and it will converge on what is the most likely output or answer for that particular problem. In fact, people get really pissed off if the AI comes back at them and throws something that they weren't wanting to do. You know, uh I don't this is just popping in my head, but I remember last year, this was a while ago, but someone posted on social media that they were asking the AI to write some code for them. And instead of writing the code, the AI responded back, telling them that they think the AI said to them, I think instead of me writing the code, you should really go off and read these things to become a better programmer, probably because what they were asking them to do was just like flat out wrong. And the person was so pissed off, and everyone responded to that social media post saying, like, oh my God, this is the end of humanity and AI is revolting against us and stuff like that. But it's like, you know, like people don't even want these systems right now to be thinking, you know, there's this risk that we don't even want them to not obey us, right? To say to us, like, hey, what you're doing is probably you haven't thought of this thing. Like, are you sure we should be doing this? Like, people would be like losing their mind if it does that. No, it should be.

SPEAKER_03 39:36

Well, I think that that's you know, you go back to like different um of these different models and like how they're how they're building them and what they're trying to achieve. Because I feel like Claude is pushy, like it's pushy. It's like, and I'm like, hey, you don't know all the facts, like let me give you some more facts, especially kind of more before we switched over uh to kind of the agentic era. But but I'll tell you, chat GBT, unless you turn off like nice guy mode or you, you know, you have a devil's advocate going on, it's just gonna tell you what it wants. Like it's and that I think that that's the approach. And people started using chat GBT and they're like, Oh, I'm always right and I'm amazing and all this kind of stuff. And then you switch over to Claude and Claude, like, you know, or Grok will tell you like what it is, and then you're like, Whoa, like I didn't expect you to say that, right?

SPEAKER_01 40:28

You know, and et cetera, yeah. I mean, it you can take the analogy also with people when they hire people, right? If I hire someone for 15 bucks an hour, 20 bucks an hour, whatever, and I tell them what to do. Most people don't want that person saying to you, like, are you sure this is the job I should be doing, right? Like they just you just want the person to do the job. Then you pay someone$200 an hour and you tell them what should be done, and they come back and say, actually, you should be thinking about these things, right? All of a sudden you really want to hear that. And if that person doesn't push back, you think that that person doesn't know what the hell they're talking about, thinking, you know. So there this also, I think, is translating across models and different price points. And the only this is gonna this is a bit controversial, but I think you could totally take giving an idea here to OpenAI, you could totally take your model, the the mass model, and just only remove one training thing, which is that when you're training the model, don't make it please the user, but instead be like more contrarian and and pushing back more, and then make people pay more money for that, and they will take that model more seriously because they're paying more money for a model to actually um you know push them rather than just please them. But you know, but most people want cheap and free things to please them, they don't want them to contradict them. So there's this psychology in that as well.

Orchestration And System Bridges

SPEAKER_03 41:46

No, it's it's super interesting. Okay, so I we have a few more minutes. I would love to talk a little bit more about kind of how you think about orchestration, because like a lot of how I was thinking about orchestration before was like, okay, one agent specializes in one skill. And now you can start to add a lot of like skills to one agent and make it like a super agent, right? But there still is different roles that you want to define uh for it to do. But how I was thinking about orchestration before, and even like some of the books that I had bought and read, like one third of the book is maybe useful now, and two thirds is like we've already passed like everything it was talking about. And so where on kind of frontier stuff, when you're setting up uh these systems, how are you thinking about orchestration today?

SPEAKER_01 42:40

Yeah, so there's two parts of orchestration. There's the human orchestration part of it, and then there's the agent-to-agent orchestration. You're talking about more the layer, right?

SPEAKER_03 42:48

Yeah, agent to agent, yeah.

SPEAKER_01 42:50

Yeah, so I would say two things. First is the architecture of it, and the second is the solution problem-driven part of it. So if you're getting inconsistent results from a process, the process you might be putting too much agency on one agent, and you need to split that up into sub-agents or sub-steps, and you might need to formalize them to become more of a linear process rather than a qualitative decision matrix, right? So, one problem I see often is one agent will be given too much too much space, too many skills, too many things it could do, and you're asking the agent to then make broad decisions instead of making it more controlled. And so, if you can, one quick solution to that would be you've got an agent which first does a decision on what type of request is coming in. Is it a, you know, uh secretarial request? Is it a booking request? Is it a research request from my human, whatever it might be, or from the clock or from some other info, whatever it is? And then it dispatches that to another agent, which then specializes on just that. And then Um that that state uh script driven element to it is also important there. Splitting it up can also help you with that memory management context. I know we probably we may not have time to get into that. I know that was one of your questions, but memory management is, of course, really important right now. Context management, context firewalling as well between agents.

SPEAKER_03 44:21

And then you talked a little bit about the bridge, like building like little bridges. Like, can you give an example of that maybe?

SPEAKER_01 44:29

So if I'm I hope I'm talking when I talk about stop me if I'm going down the wrong path here with the bridge, you're talking going over the wrong bridge for what you're talking about. But I would are you talking about like the, for example, that SEOGO agent who comes up with a brief and then the connection to the next agent. Yeah, yeah.

SPEAKER_03 44:43

Like you have to build a little of the brief to do the bigger project, like giving people an example of what that is, you you can start to separate out what's being done, and you're connecting like maybe two systems together, if you will.

SPEAKER_01 44:56

Yeah. So going back to my analogy of thinking about jobs. So if you think about a job and the domain of what that job is and the knowledge that's needed for that, that gives you a contained space, that job may sometimes need to be split up if you're putting too many jobs into it. You know, I sometimes I'll look at a job description for a human, and then when I look at it for an AI, it's actually three jobs because it makes more sense that way. They're just trying to hire some kind of unicorn to do these multiple roles all together. And then, but then just like in in traditional systems, the some of the most important parts is the connectors, the edges between systems, those bridges, those gaps, or APIs, or whatever you want to call about it. And so you need to formalize the output to the input of the next system and how you're going to control for handover and errors and expectations and checks between these systems. One sec, I have to clear my throat for a second. Okay, sorry about that. And my and my coffee cup's empty, so I couldn't just liquid liquidate my throat there. Liquidate, liquefy, I don't know, whatever the word is, lubricate. That's what I'm looking for. Um anyway, so you have you have to look at that bridge point, and then there's different ways you could do that. You could use an inbox system where you just talk to each other, you could use a folder-based system where they scan those folders, you could use uh more like a spreadsheet, right? Like you drop it on a spreadsheet and it could be a database. I don't know, it depends on how sophisticated it needs to be, how much you need to scale it.

SPEAKER_03 46:23

Because I I I was like building stuff that I was having perplexity create, and I wanted it to talk to Claud Code. And the best way I could figure out how to do it is dump everything into a spreadsheet, and then Claude Code recalled it. Like I was trying to get them to talk to each other, uh, but I just had to like drop everything um in a document and then you know have it uh when you say complexity, it's just you using perplexity browser and so comet, uh, you can spin up multiple agents at a time to do like different kinds of research.

SPEAKER_01 46:57

Okay, yeah, yeah, yeah.

SPEAKER_03 46:58

And so so, but I was trying to get them to talk to each other, and I I was just dropping in a spreadsheet and then recalling when it was updated so it would check it, and if new information was checked, it would pull it. Uh, it's kind of how I got it set up, but it's pretty rude for it.

SPEAKER_01 47:13

When we're using yeah, when we're using perplexity, because it's great for certain, it's really good for certain things. We're using it as an API tool for an agent that itself has access to databases or memory or whatever. So we don't need to that whole kind of what you're talking about, we don't end up having to deal with it because agent, I don't know, Bob. That's called agent Bob. But Bob has the API perplexity, does research with there, gets data back, stores it either in a JSON file or a table or superbase or whatever kind of needed. And then the next agent can just pull straight from that system. So it's just all within the same ecosystem. Yeah, not having to translate across a third medium. Got it.

SPEAKER_03 47:53

Yeah, yeah, awesome. So I I I do have a hard stop on the hour. I would love to kind of I know I I jumped around. I'll I'll I'll definitely have to have you back. I I really enjoyed our conversation. I know we were talking some philosophical stuff. Um, yeah, no, you try to talk to somebody on the street about this. Sometimes uh people are like, what are you talking about? Oh, totally. Um, no, maybe uh is there anything in the course of this conversation that you thought might be really valuable to just add to to this conversation here? And then if you could share a little bit more about what uh Fountain City is doing and how to get in touch with you. Uh, you talked about some blogs you're writing, uh, and that would be phenomenal.

SPEAKER_01 48:40

Things that weren't discussed. I mean, there's so many things that could be talked about, you know, the security and how are you handling that?

SPEAKER_03 48:46

There's scalability, there's don't drop your don't drop your API keys right in the prop, guys. Like, don't do that because it's gonna get processed. Like there, you gotta you gotta set up a file or something like that where it pulls from.

Fountain City Projects And Where To Follow

SPEAKER_01 48:58

So yeah. Yeah. So there's there's a lot to talk about there. Um I I yeah, I think every everything I can think of is just rabbit holes. So I'm I won't bring up like too many things that haven't been talked about. But in terms of me, us, fountain city, so fountaincity.tech is the website, and our clients right now are actually a lot of agencies are hiring us to build either systems for themselves and or for their clients. Sometimes uh one person also we're talking to right now is to build a vertical stack system. So it's they have lots of really tiny clients, like 80 clients in one vertical. Uh, I won't name it, but you know, imagine it could be like, I don't know, plumbing or services, yeah.

SPEAKER_02 49:45

So we've got a service.

SPEAKER_01 49:45

So some kind of services industry or some kind of industry. And so then you you get this economy scale because you can have one research pipeline, and then it generates a whole bunch of topics, which then get interpreted and resegmented for different writers for each um of the different subdomains. So we're doing a lot, yeah, like I said, with agencies, direct to client, and then these these uh integrated verticals or multi-site hosted stuff. So that's kind of one part of our business. And then the other part is doing what I described with the agenda coding, because we're able to, and then all those projects tend to be building systems that have AI in it. So, like a uh right now we've got two biggest projects we're running is one is a voice system where it listens into enterprise conversations and then gives very and then listens into these enterprise leaders speaking and then gives very thoughtful input to kind of help them to think broader or think deeper or think in a new way about the problems they're bringing up. It's a very interesting project. And then the other one we're doing right now is on a data intelligence layer. So we're connecting together. People come to the website from all different possible paths, everything they interact on the website, and then everything they've purchased or thinking about as they touch all these different systems, and then building that out into on-demand charts, uh, being able to predict, you know, who should they go talk to first, and then being able to AI talk to the data and stuff like that. So that's another cool kind of project we're doing.

SPEAKER_03 51:10

So I'm using a tool called uh E-R A F A N A. Are you familiar? I don't even know how to say that.

SPEAKER_01 51:18

Can you pronounce that? Sorry, Grafna? Is that what you mean?

SPEAKER_03 51:21

Grafna, yeah, Grafna. Um, basically to to to take like that data intelligence layer that you're talking about, yeah, to throw it into these kind of dashboards. You can look it up. You would I don't know what you're using. I would love to have a tools conversation with you.

SPEAKER_01 51:36

Um we're using grist on that one, okay, which is another G letter. Yeah. It's not grapna, it's grist. It's close. I was almost wondering if you're talking about that. But yeah, we are using that right now as kind of a it's it's kind of like a it's like a spreadsheet in a database had a baby kind of thing. So it's a bit like a spread, or it's like a almost like a uh what was that thing called? There was the spreadsheet thing that was popular with project managers a few years ago, where it's like super spreadsheets. So it's kind of like that because they this client really likes spreadsheets, so we're giving them that as like a low-cost kind of middleware to all of the data collection, and then from there, because it's essentially database, we can connect it to anything we want, visualization, AI model context and stuff like that.

SPEAKER_03 52:23

Yeah, this would be like the next generation of like BI or looker or something like that that's more interactive and you can pull in more data. No, I uh Sebastian, this conversation's been been awesome. So, where are you putting out new new information? You talked about your blog. Is there like do you have socials stuff?

SPEAKER_01 52:43

Yeah, so LinkedIn, my account's probably the best. Uh so just my name on LinkedIn and um YouTube. I think it's been a bit slower on YouTube recently, but I I it used to be once a week, now it's kind of like once a month. But I'll probably pick it back up a little bit more. I just have some speaking events coming up, so I've been and this launch of this kind of virtual digital agency product has been very time consuming. So I've been more focused on that recently, this kind of autonomous agent system. But anyway, YouTube, LinkedIn, the blog on our website is almost daily blog posts on everything related to this, like you know, from security to uh implementate implement implementation. I'm stumbling over my words, and then um what else? There is a newsletter also on Fountain City. It's pretty, it's not very often that we newsletter out, so you won't get overloaded, but you can sign up for that too on our website.

SPEAKER_03 53:38

Awesome. Well, we'll definitely have to have you back on. Um, I I have so many more questions for you, and and uh I like all the stuff you're doing. If anybody's listening, this is kind of where where it's all going. Um, and so you need to I I think really understand the um taste. Is that what we said? Is that what we said on like the taste of kind of where you what is your secret sauce? Yeah, like why people should hire you. What are what are you doing? I think that that pin pendulum is going to kind of swing back because a lot of these things are gonna be compressed and um you know, uh programmatic is just, you know, like the world's changing. Um and so uh it's hard to stay on top of all this stuff. Um, Sebastian, it's great to have you on. Um, thank you for being here, guys. Go check out what he's doing. Um, I think uh, you know, this is the future. Um until the next time, if you want to grow your business with the largest, most powerful tool on the planet, which I would have said is the internet, but you know, I would say it's now agentic agents. Uh keep listening to this podcast. We'll keep bringing on um some of the people in the industry that are moving it forward. Uh, thank you so much for your time. Uh until next time. My name is Bat Bertram. Bye bye for now.