The Leanpub Podcast 🎙 Feat. Ediz Najim, Author of Systems Thinking for Agentic AI

Len: Hi, I’m Len Epp, co-founder of Leanpub, and in this episode of the Leanpub Podcast, I’ll be interviewing Ediz Najim. Based in Istanbul, Ediz is a software architect and technical leader with nearly two decades of experience designing and building distributed systems, microservices, and production-grade backend platforms. His work focuses on event-driven architecture, reliability, performance, and the practical engineering of AI-enabled software systems.

Ediz is the author of the Leanpub book, Systems Thinking for Agentic AI: A Software Architect’s Guide to Building Reliable LLM and Agent Systems. In the book, Ediz shows software engineers and architects how to design reliable AI applications with prompts, RAG, tools, memory, orchestration, guardrails, evaluation, observability, and runtime control. The book explains how to move from simple chatbot experiments to reliable AI-enabled software systems — reliable being a keyword there. It focuses on the production realities that matter after the demo works — after the demo works being another key thing — latency, cost, failure handling, tool execution, structured outputs, testing, tracing, safety controls, and maintainability.

In this interview, we’re going to talk about Ediz’s background and career, his book, and at the end, we’ll talk a little bit about his approach to writing and being an author. So thank you very much, Ediz, for being on the Leanpub Podcast.

Ediz: First of all, thanks for having me. I’m glad to meet you again on the podcast.

Len: Yeah, thanks. We did a launch video previously. Let me just pop up the cover of the book here. OK, so here’s the book we’re going to be talking about in a bit. But first, I would like to ask you about yourself. I always like to start these podcast interviews by asking people for their origin story. So I was wondering if you could talk a little bit about yourself and where you grew up and how you built your career and how you first became interested in computers and technology.

Ediz: Sure. I was born in Baghdad. I lived with a big family. They encouraged me, supported me to do my best. And, to be honest, I was a very good student. I loved to learn, searching and reading books. When I finished my university, or after finishing my university, I moved to Istanbul and started my professional career in Istanbul in a big company. And my first project was very challenging for me because it was written in Java and in the same framework. And, you know, at that time, I didn’t have good information or knowledge to manage this project. But I worked a lot — and it is normal to work at midnight or wake up earlier and look at the project and the classes. And after two months, my IT manager told me that I must be responsible for the mobile web too, and that let me be confident in myself and feel good, because a big company giving me more responsibilities means I did very well. And I started to search more, read more, and write articles — at that time, in 2008.

And after that, I decided to work in startup projects, because I believe that startup projects will teach me a lot of new information and a lot of new technologies. For that, I started to work with small companies, and I got the freedom of choosing the technology and the frameworks. It was very good for me to work with these small companies. And my dream was that one day I would work as a software architect, and that let me be really working all the time without stopping, to achieve my dream.

Len: OK, fantastic. Thank you for that great answer. One question I was looking forward to asking you about was studying at university in Iraq. I went to university in Canada and in the United Kingdom, and I’ve interviewed authors from all around the world — we’ve learned about Germany and Spain and Italy and China. But you’re the first guest I think we’ve had where I get to ask you: what’s the university system like in Iraq? Is it four-year degrees? Is it three-year degrees for undergrad? What is it like?

Ediz: Yes, we have a good education system in Iraq. The education system is not different from other countries. And if you search about the good universities in Iraq, you will find a lot of researchers and thinkers who work there. For that, I am thankful to study in Baghdad and get my degree from Baghdad. But, you know, the university is different from working. For that, the university gives us the basics. But at the end of the day, we must develop ourselves. We must search. We must learn a lot of things from our mistakes, especially when we are working in the software field.

Len: Yeah, that’s a great answer, because it leads me to my next question, which I often ask authors with a background like yours, which is: if you were starting out now to have a career as a software architect, would you go to university?

Ediz: Yes. Because the academic information is really important — important to get this knowledge from academic people who will teach you how to search, how to be clever. I believe that all people are clever, but most of them don’t use their brain. I’m sorry about that. But if a human uses the brain in a good manner, he can achieve everything. We still need universities. But I have one point about universities: we must not give only basic information. We must let the students attend labs and attend — how can I say that — real-world software, building real-world software.

Len: Yeah, I’ve heard lots of different answers to that question. I very much agree with the answer that you gave myself. One thing — the way some university programs work in Canada is that when someone’s studying software engineering, they have to do what’s called co-op terms. So they have to go work for a company three or four times during their degree. And we have, I think, three or four co-op students at Leanpub right now. And it’s just amazing seeing them learn day to day from the real-world experience and then go back to their classes with that experience. It’s a very good model.

Ediz: Let me explain something here, please. We have this system, but it is not enough. I want to explain that to you. It is not enough to go one month or two months to a company and work with their employees. We must make the practical like academic information. That is what I want to say. It is not enough to go one or two months to the companies. It must be more.

Len: OK, yes. Yeah, that makes sense. So moving on to talk about your book — which I have up here on the video, for the second time in the history of the Leanpub video sort of things, I just learned how to do this — the book is very good and it’s very thorough and systematic. And although we’ve talked about it already on a launch video, I just wanted to ask you a little bit about the origin story of the book. When did you start writing the book? Did it start as blog posts, things like that?

Ediz: Sure, sure. As I mentioned, I like to write everything, or write notes in drafts. And I like to post on LinkedIn or other platforms. This led me to collect these drafts and put them together. And in the end of the day, I saw that I could write a book. The idea of this book was not about agentic AI — it was about a Java framework many years ago. But, you know, we are busy living and achieving some goals in our life. This time is the right time to write this book, to put my experience in software architecture together with my work on agentic AI, and give developers, software engineers, and tech leads one book. One book that contains all they need to build a reliable, traceable, and secure agentic AI system. For that, my book is not speaking about the basics or diving into the details. No, I’m speaking about the infrastructure, about the architecture, and how we can apply systems thinking to agentic AI. For that, if you don’t have any information about AI, you can read my book. If you are a professional in AI, you can read my book too.

Len: Yeah, and I love the idea of systems thinking and emphasizing the thinking part. And there’s a sort of theoretical foundation to your book as well. So my first question about that is: what do you mean when you say an LLM is not an AI system?

Ediz: Yes, many teams have a misunderstanding about that, because they still think the LLM is the agentic AI. But in the real world — in fact, in the software world — the LLM is only one component in an agentic AI system. You remember when I gave you an example: an LLM is like our brain. Our brain can listen, think, make decisions. But in the end of the day, we need to use our hands or our eyes or whatever to achieve what we want to do. Exactly. An LLM can think, can decide, can do something. But in the end of the day, an LLM needs other surrounding components to do what we want from it as a system, not only as an LLM or only as AI.

Len: And so, yes — when you’re talking about building a system, you’re talking about essentially, metaphorically, building a body that has interacting parts. Some parts can do this, some other parts can do that. And you get to decide when you’re building the system what to build.

Ediz: Exactly, exactly, exactly. You cannot make the LLM the endpoint of permission or authorization. We must validate what comes from the user. We must give the LLM the tools. We must give the LLM specific information about the user’s question. And in the end of the process, we must validate the output of the LLM, because the LLM can give us fluent answers but wrong answers. For that, we must have rules to apply over the LLM results and see if we have the right evidence to say, yes, the result is OK for the end user. For that, all the surrounding components together with the LLM will create an agentic AI system. Only the LLM alone will be only AI, or only one component.

Len: Yeah, it’s funny — one of the reasons I love your brain-body kind of metaphor is I have this joke that we achieved artificial general intelligence when they started hallucinating and making mistakes, because that’s exactly what we do all the time. But so it’s very interesting, this idea of taking something that’s non-deterministic, very powerful, and then building a system around it so it can be used in various ways, but very crucially, reliably. And you use that word “reliable” and “reliability” in the description of your book that I gave earlier. So I was wondering if you could talk a little bit about what you mean when you talk in the book about reliability. What does that really mean when you’re building a system?

Ediz: As you mentioned, the LLM can answer us with wrong information, and that will lead us to wrong results. And it does not matter when you ask the LLM to generate a paragraph or to edit an article. But it is very important when you ask the LLM to make a payment, or accept a payment, or maybe control a bank account. For that, if you say “don’t accept a payment of one thousand dollars,” this is only the instruction, but you will not enforce the LLM to apply this correctly. For that, we need reliable components to validate the input and the output. And we need the pipeline to record every step, which will let us trace the system and observe the system.

Can you imagine that when one day a customer calls you and says, “one million dollars was paid from our system” — you must find where this money went and why we made this mistake. For that, reliability is very important. We are not working here in a deterministic system. A deterministic system means one plus one equals two, if-else. Now we are working with — let me say — and I’m used to saying that an LLM is like a child. We need to speak with it phrase by phrase, step by step, and the prompts must be clear.

When you ask — I’m a father of two sons, Alp and Amin, and one daughter, Alina, and I know very well that when I am speaking with my children, I try to give them instructions one by one. This is what we must do with LLMs. Otherwise, yes, we will get an unreliable system and we will live with many troubles.

Len: Yeah, I really love that idea. Particularly comparing it to children, but not comparing it — saying it actually is like that. And particularly setting things out step by step is really important. I think that’s something I’ll say I’m guilty of not doing myself when I work with Claude or something like that. And as soon as you run into trouble, you realize you have no idea what’s going on between the input and the output, unless you then have some concept of the steps that are being carried out in between. And so, you can’t — it’s basically just guessing and hoping until you’ve actually worked out a system step by step and understand what’s happening at each stage along the way.

So, let’s say you’ve built a system — an agentic AI system around an LLM — how do you evaluate whether it’s working well?

Ediz: Yes, this is one of the biggest problems that we must solve in an agentic AI world. I explained this in my book and I showed how important evaluation is. If you — or anyone — cannot evaluate the system, then you will not know if the system is working well or badly. For that, evaluation is very important in all software. Here in agentic AI, because we have a non-deterministic component, this non-deterministic component can generate anything. Maybe you will say “hello” and it says “you can’t drink coffee.” What is the relation between “hello” and “you can’t drink coffee”? You must evaluate the results and the inputs. Evaluation is very important, as I mentioned, and we don’t need another framework or another tool to make this evaluation. All we need is to know the input, validate the input according to some rules, and know what evidence can tell us whether the result is OK or not.

Len: Right. So you have to set out the evidence for the evaluation and think about that. That’s very interesting. One thing you reminded me of was — I think an author I was doing a launch video with recently who was saying, look, when it comes to evaluation in the old way of doing things, when it was deterministic — for 80 years of computing, you built a system, you evaluated it by running it and then you walked away, like, it works. But that’s not how evaluation works with agentic AI.

Ediz: Yes, yes, it is different. When we come to evaluation, it is different in agentic AI. Because as I told you, one plus one equals two — that is the rule. We can make unit tests for testing the system and everything will be OK. But now we have a component that can produce anything — maybe good, maybe bad, maybe related, maybe unrelated.

Len: So I’m going to quote you and ask you a question, which is sometimes an unfair thing to do as an interviewer, but we’ll see. You have a very good line in your book that really distills things well. You say, in your discussion of agents: “An agent is a system that uses a model inside a controlled loop to pursue a task through actions, observations, and state until a stopping condition is reached.” Now my question to you is about the “controlled loop” part. How is the loop controlled?

Ediz: Yes. In the last few years, we always read about loop engineering. Loop engineering is meant to make a loop that controls everything before the result. The same idea applies here when we write agentic AI — we need a loop, and we must be aware of some things. I will mention them. The loop will take the input of the user and make validation and make some steps. And after that, it will ask the LLM to process this input. The LLM may need to execute a tool. The LLM will ask the loop control to execute this method and return back the result. And this loop will continue until the LLM says the final result is ready.

I say we must be aware about some items. One of them: this loop must not be an infinite loop. We must have a condition to stop this loop. Maybe we will say: try only ten times — ten times you will go to the LLM. Or we can say you will consume only one thousand tokens. For that, yes, loop engineering is very important in an agentic AI world. It’s very important to make loop control to control the input and the result and be the layer between the user and the LLM.

Len: That’s a very, very clear answer. It really helped me understand better. My next question is — and I’m going to be quoting you back at yourself again. Why does an application need — and here I’m quoting you — “an explicit registry that defines which tools exist, what each tool is for, what input schema it accepts, and which execution handler or back-end operation it maps to”? So you don’t need to obviously remember all those specific things, but what is the need for this explicit registry of tools?

Ediz: Really, for these tools, we need the MCP — Model Context Protocol — to keep all these tools or methods with signatures that say: this tool can accept these parameters and return this type of value. And if we put all these tools in one place — MCP — we will focus on another part: the part to connect with the external world. Like, I want to get the weather conditions in Istanbul or in London. Then I, as the LLM, will not be worried about the web service or about the URL. I will only say, “I need the weather in Istanbul,” and it is enough. MCP will go to execute a URL, maybe, or get this information from a database, maybe use Redis or another technology to get this information.

Len: My last question about the book is a hard one, I think. What do you think the role of an enterprise software architect is going to look like in five years? Will it be the same as it is now? Will it be very different?

Ediz: Yes, nowadays we are starting to speak about agents, about AI generally — you know, GPT, Claude, Gemini, and other tools — and every day, I can say, we see a new version. For that, I think the software world will be changed. AI will be used everywhere. But we need to be aware about AI. We must not give everything to this AI to do, because the maturity of AI is not that good now. But five years from now, maybe we will see a component or a system that lets us say, yes, I can trust that AI may change all the software world.

Len: That’s a really great answer. And that leads me to the next part of the interview, where we talk about your work as a writer. You’ve talked a little bit about this already. But one question I always like to ask people, because writing a book is a big project — how did you approach it? In writing generally, whether it’s LinkedIn posts, etc., do you set aside time — like, do you put a thing in your calendar and say from two to three o’clock tomorrow afternoon I’m going to be writing? Do you go to cafes? Do you like to write on the train? Do you have a study where you tell your kids they can’t come in if you’re in there and the door is closed? How do you do that?

Ediz: Yes. The people who know me will answer this question easily, because I’m working all the time. Yes, I’m working all the time. I don’t have a schedule to finish my work and take a rest. And every day, my goal is to write one or two articles about agentic AI generally. That lets me be creative in writing articles and finding other things. Maybe you saw my STAAI architecture model — a model that I’m trying to build to be the standard for building agentic AI, because I didn’t find standards or principles to write agentic AI. For that, you can visit staai.com to get all the information about the STAAI architecture model and about the STAAI principles that I’m working on.

Len: Yes, so just so people know, it’s STAAI.com — is that correct?

Ediz: STAAI.com, or staaicore.com.

Len: OK, sorry, please go on. I just wanted to make sure people could find it.

Ediz: Yes, STAAI.com.

Len: OK. That’s awesome. What a great project. And so the last question I always like to ask people if they’ve written a book on Leanpub — or they’re using Leanpub to publish their book — is: if there was one magical feature we could build for you that you could ask us to build, or if there’s one thing you totally hate about Leanpub and you think is terrible that you could ask us to fix, can you think of anything you would ask us to do?

Ediz: I couldn’t understand you — where do you mean? Because the internet is not that stable.

Len: Oh, OK. So my last question for the guest is usually about Leanpub. And it’s: if there was one feature we could build for you — anything you could ask us to do — can you think of anything you would ask us to do?

Ediz: Really, your platform is very good and very nice. And you take care about the authors. You meet the authors, ask them. And really, it is very good for us to do this podcast or interview. One thing — maybe print books. This will be great for us to say we have a print book on Leanpub.

Len: Yeah, thank you very much for that. We’ve had that request from many authors over the years. You know, the world of paper things and the world of bits on screens is a very different kind of business. And so we don’t do print books. But what we do is we try to make it as easy as possible for people. So if you write a book in a Leanpub writing mode, we have a print-ready PDF that you can get. We do some settings and you do one click, and then you can take that and you can go to Amazon KDP or IngramSpark or other things and you can do it yourself. You can also pay us — if you want to pay us a lot of money — we can go and manage all of that for you on Amazon KDP. But the main thing is we try to make it easy: click a button, get the PDF you need, and then you can go to a print-on-demand service and get print. But yes, thank you. That’s basically the hardest thing you could ask us to do, and many other authors have asked us to do it.

Ediz: I wait for that service from you, so that I don’t need to go to another platform or website.

Len: So just before we go, is there anyone you would like to thank who may have helped in the process of writing the book, or in developing your career and things like that?

Ediz: This is a very good question. Really, I want to thank my mother and my father, who encouraged me and made me what I am now, because without them, believe me, I will not be here. I will not be sitting down and speaking face to face with you. For that, thank you. Thanks, Mom. Thanks for everything.

Len: Well, Ediz, thank you so much for doing a video with me again. This time we did a full podcast instead of just a launch video. It’s just great — always great to talk to you and to hear your thoughts. And thank you for using Leanpub to publish your great book. It’s Systems Thinking for Agentic AI: A Software Architect’s Guide to Building Reliable LLM and Agent Systems. So thank you so much for being on the Leanpub Podcast.

Ediz: Thank you very much for your time. I’m glad to meet you. Thank you for having me. Thank you.