Sign in or Join the community to continue
Everyone Uses AI Agents. So Where Are the Agent Builders?
Posted Nov 10, 2026 | Views 6
Share
Speakers
Erica Hughberg
Community Advocate @ Tetrate
Erica Hughberg is a technical leader and community advocate passionate about helping engineering teams build scalable, secure, and human-centric application platforms. With a background in software engineering and a deep understanding of cloud-native technologies, she specializes in driving the adoption of open-source projects like Envoy Gateway, Istio, and Kubernetes Gateway API, which enable organizations to simplify traffic management, security, and API distribution.
As a maintainer of Envoy AI Gateway, she plays a key role in shaping the future of API infrastructure. She focuses on features to ensure organizations can securely and efficiently integrate AI-powered services while simplifying traffic management, security, and API distribution. In the Envoy community, she drives collaboration, mentorship, and contributions that advance the project and its adoption.
Lastly, as a believer in the power of storytelling, Erica enjoys translating complex technical concepts into engaging, accessible narratives in the form of social media posts, conference talks, podcasts, and educational content.
+ Read More
Demetrios Brinkmann
Chief Happiness Engineer @ MLOps Community
At the moment Demetrios is immersing himself in Machine Learning by interviewing experts from around the world in the weekly MLOps.community meetups. Demetrios is constantly learning and engaging in new activities to get uncomfortable and learn from his mistakes. He tries to bring creativity into every aspect of his life, whether that be analyzing the best paths forward, overcoming obstacles, or building lego houses with his daughter.
+ Read More
SUMMARY
Lots of companies say they have AI agents in production. Far fewer seem to have built their own, and almost nobody wants to admit it's really hard.
This conversation pulls apart what "agents in production" actually means, with both speakers drawing on their own work: one is hosting a virtual conference for the third time, the other is prepping a livestream for AgentCon and MCPCon in San Jose. Their case: plenty of agents already do real work. They're just Claude Code and Codex, not homegrown platforms, and nobody earns brownie points for saying so.
+ Read More
CONTENT & TRANSCRIPT
Demetrios (00:00) Is it just posturing? Like, are we just seeing a lot of people talk about how they have agents in their companies just because they can't be the person who says, yeah, actually, we're trying this and it's really hard.
Erica Hughberg (00:17) I have been thinking about this. It's not the models that's holding us back. It's the systems around the models. I feel that the Agentic AI push comes from boardrooms. It comes from leaders. Not from we have a problem we have to solve. It comes from what is our AI or agent strategy. Gartner predicts 70%, seven zero of enterprises will abandon Agenetic AI built by vendor forward deployed engineering by 2028.
Erica Hughberg (01:01) I have started to feel like. There is a lot of people who want to build agents or they wish they did, but there's not a lot of companies that are actually putting them into production. So I got a bit interested in this phenomenon, if you will. I also, and it sort of also relates to this entire topic of MCP and security and security in general. I feel like the hump into production. is kind of big or like into trusted environments or like business critical decisions.
Demetrios (01:42) And so first thing that comes to my mind is when you say agents in production, what do you mean by that?
Erica Hughberg (01:50) Okay. I mean that someone has actually built a, like they're not using another company's agents. Like I feel like when people say they use code code or codex or other things like that, it's you're using an agent from a provider. Okay. That's cool. You might be using that in a production or business critical workflow. What I'm curious about is. Where are the agent builders, the people that really build agents and deploy them and run them beyond like the big labs and some of the very well known, like you have the Lovable, the Legoros, the, you know, other big companies that have agents that people use, Replit, whatever. Right. Then, so I thought about that and. I have a feeling and you know, when you have feelings and you don't have like proper data to back it up. So I've been, I feel when I speak to people that there isn't a lot of, we built a fleet of agents that we're running that are taking business critical decisions in production. So when you say like, what do I mean by production? Imagine. I even think there's a lot of SaaS providers that could build agents for their users that they can use. So for example, I use Netlify, by the way, and Netlify now has a type of agent that can do things for you in Netlify. So for those who don't know what Netlify is, it's a place where you can deploy your websites and stuff like that. And you can have very lightweight functions running. So you can build a very simple website, if you will. And yeah, you can get like an agent to help you build a website and you can get an agent to even help you understand where your deployment might have failed. So they're like built in. So I would say that that is Netlify running agents in production, but that's for their user, right? It's not agents for, maybe Netlify has agents in production for their own workflows and their own work. But I just feel like the amount of companies that have truly have. Agents at scale in production, doing work. I'm struggling to find it. I mean, you also see like Bloomberg has agents in production they have agentic features again for the clients that are using Bloomberg right
Demetrios (04:28) In the boom
Erica Hughberg (04:30) Yeah, I don't know exactly where it is because I am not a fancy person in finance anymore using Bloomberg but they have you know according to their press releases and marketing right they have agentic uh features for their users
Demetrios (04:49) Yeah, I feel, there's so much to parse uh in this which is fascinating that you're bringing it all up right the first thing is that ramp came out with their canonical blog post I think probably a year ago now in 2025 at some point that said here's how we're using agents here's our agent building platform at ramp and everyone is getting so much more value out of their job and being way more efficient because they can build agents on our agent platform And so then I think that became the canonical standard. Like, whoa, Ramp can do this. They have this platform that makes things much better. Okay, let's see. But then when you really dig in and you go, okay, then is it just posturing? Like, are we just seeing a lot of people talk about how they have agents in their companies just because they can't be the person who says, yeah, actually, we're trying this and it's really hard. We haven't really figured it out. And you know what? We created a platform, but it hasn't worked out like we thought it was going to work out. You're never going to see anybody say that because that just like it admits defeat, right?
Erica Hughberg (06:13) Yeah. I think that's also like admitting defeat or admitting that you don't need it in the way you may be envisioned. So, I was thinking about back in like 2015, 16, for those who were around in software, there was a big boom around microservices. So we were moving from this notion of monoliths into microservices. So there was a huge push to move to microservices because we all needed to scale, you see. We all needed to be able to scale and we needed to adopt Kubernetes when that came out a couple of years later, right? So first we came into microservices, then we got Kubernetes to help us manage that. And like by 2019, 18, we all needed to start using Kubernetes, right? And be cloud native because we all needed to scale. What to me is interestingly different from that era versus the AI agentic era is that the, sorry, I'm coughing, that what I think is interesting is compared to that time in engineering evolution or revolution, if you will, the push came from a different point. I feel that the agentic AI push comes from boardrooms. It comes from leaders like. Not from we have a problem we have to solve. It comes from what is our AI or Agent strategy. And then it comes down towards the builders.
Demetrios (07:44) But you also have the bottom up because folks are using Cloud Code or Codex every day. And so for that use case, I think people are seeing, oh my gosh, there's so much value. And that's why that's kind of the other thing. exactly I wanted to parse out this other piece when we're talking about these agents in production yeah because I almost would argue that yes there are a lot of agents in production it's just through cloud code and codex and so considering how much usage so many companies are getting out of that those are technically agents in production it's just maybe on the back end and it's helping them in that way and it's not like I built my own agent or I had to create my own harness and then we have now productionized that so that the whole company can use this one particular agent which is more like a workflow but we call it an agent.
Erica Hughberg (08:42) Yeah like inference enhanced workflow is your now agent. But yes, I think this is like an important distinction that I've been thinking about. It's like using agents that are vendor built, right? Vendor provided. I think it's a much, yeah, that's a huge adoption of that. And yeah, you're right from that perspective, that's been very engineering driven, but I don't think that. when you're looking for like increasing your own value of your own company you don't go out and be like hey, we use cloud code right you want to have your story of like what you are building how you're innovating
Demetrios (09:22) Yeah, because there's no brownie points for being like, we're one of Anthropic's biggest customers. You don't get the kind of fame online if you write the blog post of like, yeah, and we just threw more money at Anthropic and it turns out that made us more productive.
Erica Hughberg (09:42) Yeah. I mean, it's a great story for Anthropic, right? But it's not going to help you so much.
Demetrios (09:47) Case study. Yeah, totally that goes on Anthropic's website not on your twitter post where you get a bunch of likes
Erica Hughberg (09:56) I have seen though the Open AI at least they have given companies like effectively little awards for being big users of Open AI like the companies get little like almost trophies for like well there was that using loads of stuff.
Demetrios (10:09) Yeah, one of their dev days they had the one trillion token award or something and they had everybody who had used up more than one trillion tokens their names were on the back on the screen as they were presenting but then this was two years ago I think and it was when I don't know if you remember there was that moment in time in our circles where that guy Soham got caught for being he was working for like eight different YC companies at the same time And so the meme that I saw going around was you had Sam Altman presenting at OpenAI Dev Day. And then you had all the names of people who had hit one trillion tokens. And somebody circled Soham's name and they're like, look who also did it. Because it was like company names, you know, and then it had like Soham because he was his own little enterprise.
Erica Hughberg (11:06) Yeah, but that's like, I think. However, you know, when we talk about the usage though of like vendor provided agents, if you will, I, just as you talked about that, I made this connection of why people are sometimes so excited about MCP because they might not be building their own agents, but they might be plugging in their own like tools. home-built tools or home-built like contextual data so they need to plug their vendor agents into that right or if you are a SaaS provider Like you are linear or you are Jira or whatever, right? Or you are GitHub. Like, of course, we know the wonderful GitHub MCP server. But I think that is where it's kind of interesting what MCP actually has a lot of attention and then obviously has the security attention on it because you want to make sure that things are secure. Well, there is. Oh, sorry. I could keep going for forever. So what were you going to say?
Demetrios (12:19) Well, when you talk about that and you also go back to. The agents in production for the end users or the customers of these different companies is one thing. But there was that great tweet by somebody who said, everyone wants me to use their agent. I don't want to use your agent. I want to use my agent on your application. Yeah. Yeah. Or your system or whatever it is. And so that kind of fits that narrative of why MCP is exciting but as an end user it is weird I will say that we are if you're clicking around in the gooey of a lot of different SaaS providers these days you get the magical stars that pop up and those instantly mean ai and now you've got an agent at your fingertips that you can use within the walled garden of this Saas product.
Erica Hughberg (13:18) And on that, like I was thinking about that this, okay, I think about a lot of weird things this morning, but on that topic, I, you know, yes, we have spoken about this. I don't know if we have, have we ever spoken about that on the podcast? Maybe we have. There's like, I want to use my agent, not your agent, but then there's agents all over. Right. So I, I was thinking about how A2A, like agent to agent protocol, like remote agents. I talk to people who are like using A2A inside of their own multi-agent like lab that they're running, say on their own computer, like they're trying stuff out. But I think that the A2A moment will be when I can have my agent talk to the agent on a remote system that I didn't build, right? yeah that's what my hope would be like.
Demetrios (14:13) Exactly like what we're talking about right where Yeah my agent then can go to the this SaaS provider and talk directly to that agent yeah and then say it the thing that gets me caught up is like oh but why is that not MCP?
Erica Hughberg (14:32) I think that's such a good question. Like when is it a tool and when is it an agent? Yeah. I, okay. I'll just do my like gut feel on it. Right. My gut feel is that a tool should be deterministic and an agent should not be. But that's why should it be two different protocols? Now you may ask, and I think that would be a very good question. Like why couldn't the tool be an agent and just by defining it as agentic, an agentic tool, I now know it's not deterministic because I don't believe all MCP tools are deterministic either because I'm sure there are a bunch of agentic feature, you know, powered tools. So, but I think it's an important, like just interface distinction for anyone. Even if we forget about MCP and A2A, like if you just had it totally like standard API that you designed, I think you have to communicate to the user, this is a deterministic, like you should, same input should always give you the same response versus nah, no, this is improvising, right? You're going to get different response.
Demetrios (15:45) Yeah. I've heard. arguments to that certain agents will have certain access and so and they will know how the tool works or the the application that it is an agent of they will know and have the access inside of the system in order to be able to do something so you think about a lot of these one of the biggest use cases that I've seen for agents in production are customer service agents yeah yes that's why companies like Decagon and Sierra i think blew up so quickly because that was a very high value use case and if you think about the companies that are creating customer service agents well you don't necessarily want that just to be an MCP because there's a lot of stuff that they do on the back end with their systems that isn't necessarily like a tool call per se It is more of an agenda that needs to happen. Yeah. And so that's where I could see the difference, I guess. It still feels very fuzzy to me. Yeah. And why you would want one versus the other.
Erica Hughberg (17:13) Yeah. I feel like I need to look at this a bit more because I think it's interesting. With agent to agent, I also feel like the idea of delegating. reaching an outcome, like my agent is no longer responsible for completing the outcome, right? I'm delegating rather than, oh, I'll take a trip back to trip down memory lane maybe, like in my life of being a manager. I like to see, I like to have identified how senior someone who was working for me based on whether I delegated them tasks, I delegated them problems, or I delegated them an outcome, or I delegated them a responsibility. And honestly, if we look at agents, are you delegating a task to the agent? Are you delegating a problem you want them to solve? Like, hey, there's a bug in this code. Can you fix it? That's giving them a problem. They have to figure out the tasks to solve the problem. You delegate them a complete like outcome being like this software has to be a reliable under load. Like it needs to be fault tolerant. Then you haven't told them there are bugs in the system, but this is the outcome. And the person or agent, maybe in this case we're coding, it has to identify there are bugs that will fail to reach the outcome, right? The next level is you need the responsibilities. You need to look after this project to meet the requirements of the users and the clients, right? Of this software. That's a responsibility delegation. Like I haven't even told you the problems or outcomes we need to reach to ensure. But these are, I would tell you, here are our clients. We need to make sure that our software fulfill their needs. Anyway, that's how I determined seniority of like people who used to work for me. And I feel like it applies to how autonomous is the agent as well.
Demetrios (19:24) Have you tried each one of those?
Erica Hughberg (19:26) On an agent?
Demetrios (19:27) Yeah.
Erica Hughberg (19:28) On humans? Yes. Not on agents. I feel like the furthest I've been able, I have not been able to get to responsibility delegation to an agent.
Demetrios (19:38) So this brings up a great point of the new rage which is the personal agents and I've been playing around with dot
Erica Hughberg (19:47) Okay, I have not tell me what have you experienced
Demetrios (19:51) It's pretty cool I guess I don't know necessarily why it is different than just the normal chat however what I've been trying to do is that delegation as far as like responsibilities. The one thing that I am not sure I like is that it's all in one chat. So you don't start new threads. It's just one continuous chat and you kind of reply to things. And there's certain use cases that it is great for and other ones that you're like, I guess it's better to just start a new thread and contain the context in this thread and for example if you want to as my friend put it he was like actually these personal agents one of the best use cases is them getting you a dinner reservation because they just like spam all of these restaurants continuously trying to get you a dinner reservation and that's it's actually pretty good at that and anything that needs follow-up it's pretty good at it will follow up pretty hard on things and then it will keep tabs on them so it's not necessary I think that's the biggest difference that I've seen in a normal chat versus the personal assistant chat is that it will continuously reference oh hey this person wrote you back about this you should do you want to respond or oh hey you said this yesterday but you still haven't done it should we do it today And so it's, it's that like continuity in trying to move the ball forward that they just probably put in the prompt or something to make it. Yeah. Hey, don't forget when there's open items, keep coming back.
Erica Hughberg (21:47) It's I do think it's interesting, you know, how. we have gotten we got so trained in you know talked about starting new contexts and starting new chats like that's what we've been sort of trained to do I have thought about that a little bit like a while back because I thought about the lovable use case, if you will. In the world of using lovable, you start a project and there's only one long continuous chat for that project, right? So you're building your website and everything. It's not like you start multiple tabs. So when you work with, say, Codex or Claude Code, you can start in your project, you can start multiple threads and work trees and you can go crazy. And I think it's interesting because if you come back to the idea of interacting with a different human, right? So even if you and I interact over Slack, over iMessage, over WhatsApp and whatever other chats we use, in a weird way, we are one continuous context between the two of us, right? Yeah. So... It's an interesting thing of how we can make agents feel more like the way you operate with another human. How can…
Demetrios (23:04) Coz, we have that one continuous context?
Erica Hughberg (23:07) Yeah, however have you ever had that situation where you have a friend and you have some conversations on like Instagram chat and then you have a completely different set of conversations in a different context on like…
Demetrios (23:19) Coz, you just send memes on Instagram and stuff and then it's like on WhatsApp you actually talk serious stuff on the phone
Erica Hughberg (23:25) Like live updates.
Demetrios (23:27) Yeah. Yeah.
Erica Hughberg (23:29) But they sort of merge into one big context because maybe when you're on the phone, you'd be like, you know that cat reel I sent you? Did you see it? So, you have like cross context, but you still weirdly have different things you talk about.
Demetrios (23:44) Yeah. Yeah. I, I do think they're doing some. really cool stuff behind the scenes with the context windows at all whether it's thoughts or it's instinct or it's muse there's a ton of them that are out right now and being able to gather the context and preserve the right things at the right time is magic yeah but at the same time what is the most dangerous for all of these personal assistants is that they really want to send you notifications so it's now like hey turn on notifications so i can ping you and as soon as I haven't turned them on because I'm just anti-notifications in general right always
Erica Hughberg (24:33) I, I don't like notifications I'm like turn them off as well i don't want to see them if I want to see you I'll come see you
Demetrios (24:42) Yeah, exactly, I'll pick up my phone and I'll look at what happened but I'm not going to let my phone dictate my life and the fact that they can send you notifications proactively about different threads that may not have been finished is potentially really cool but also potentially very dangerous yeah because now it's going to be notifying you about stuff and you're like the attention yeah I don't I don't care about that or just because I never responded does that means that like it wasn't that big of a deal or I didn't explicitly say hey this item is now closed it tends to look through things and see but it could be a little bit annoying and so you have to dial down the notifications or you dial them up if you want more updates frequently about it but the whole reason I got into this was because I asked this morning as I'm doing a live stream for AgentCon and MCPCon in San Jose yeah at the event and so I told Dot I was like hey I need to start promoting that and get people to sign up for that event And it's like, I got it. I'll take care of it. And it drafted me a LinkedIn post that I could have. And then it was like, and we'll send this here and there. But it still isn't that good because it doesn't recognize all the stuff I did before. For example, the marketing that I did for the Amsterdam event. And it's weird because it kind of sees half of it from what I would interact with in chat for that event but it doesn't see the other stuff like all of the asks that I had for people in Slack or for the email newsletter that went out or for this and that. And so despite me trying to connect all of my context and to constantly context dump, it still doesn't get it. And that just shows you how horrible the problem is.
Erica Hughberg (26:59) Like, okay. On that, like, yes, I was on my weird musings of thoughts. I have been thinking about this. It's not the models that's holding us back. It's the systems around the models, right? And I feel like that's what you're touching on. Like, how can we get the right context in? How can we get the right, like, the model, the models are really good, but we aren't having the... and some harnesses are amazing, especially they've been so tuned for coding, which I think is really wonderful in a fun way that we have now amazing like coding agents allowing us to build things. And the part though is that I have ran into the same frustrations as you mentioned here. Like I'm hosting a virtual conference next week. Yes, it's next week. And it's like, again, I was doing a similar thing. Like I have branding, I have all of these things, but what's supposed to look like, et cetera. I, this is like the third time I'm running this virtual conference. So there's a process and there's. The first time we did it, we didn't use much AI at all, right? So there's like literally like a to-do list in Google Docs. And I'm like, here's my old to-do list and here is my... But then, yeah, there's those weird little things of channels in Slack and coordination and setting this and that up. And yeah, you don't get it all in there. And the amount of times I've had to reintroduce the design system, which has not changed its location. I'm like...
Demetrios (28:41) That's what it's because of the context compaction right? That's got to be…
Erica Hughberg (28:47) I don't know exactly I've been this this is me using co-work with Claude I've been trying to like have a project in co-work to see if it can like help me through it and I'm like, and I don't know if you use co-work in the projects, you can like add all the context into this project, but somehow it still seemed to be missing stuff between and back to like the continuous chat, weirdly, sometimes I felt it's easier to go back to an older chat than starting a new one in the project, because at least then we have the chat history. Which somehow makes it happier.
Demetrios (29:24) Exactly! once you get too much context again it's like you're overloading it and now you're getting worse results and so there is a really it's an it's an art there that you have to walk this very fine line of yeah what am I trying to get done and did I already talk about this do I already have a document about this it's almost better to like you said like re-reference the document or re-reference the design file each time you need to because you can't assume it's going to find it if it has all these million other documents
Erica Hughberg (30:08) Like that if you had a human that you had shared all of this with as you onboarded them to the work you would start getting really frustrated if you had to show them where the design system was.
Demetrios (30:23) So I will say I am horrible at this because I know that working with my team a lot of times, I'm like, where's that document again? Where does that live? Or can you remind me the link? Or when you're in a call, it's like drop in the link. I don't know where it is, even though it's literally on the calendar invite.
Erica Hughberg (30:43) Or it's in the Slack. Yeah.
Demetrios (30:46) Or it's in the Slack channel or it's in the email. is it in the chat of the call where is it exactly and so you really it's frustrating in that regard too
Erica Hughberg (32:59) Yeah, but imagine you had like so I, I work in a wider team where I work I work and we have all of the companies design assets in a google drive right And it's these things that I know it exists. And sometimes I don't have the link saved, but I know I can go to Google Drive and I can search for like logo assets and they will find me. So, I don't have to ask a human, right? Because I find my own way there. Like I have my little index in my head that in Google Drive, somewhere there's a folder and I can find it. But at one point a human had to tell me about this, right? They had to let me know. So, they put it into my context of when I do, say, design work. That when I do design work, I should go and use the proper files. Anyway, it's just like interesting because I get frustrated because I know that a human, if you give a human this information and they have a job to do, they're going to like figure it out to a degree, right? Well,
Demetrios (32:11) One thing I've done with my agent is I'm using this agent librarian. I think that's what it is. It's some open-source project and it downloads an embedding model onto your computer, a pretty small one, and then it chunks up all your documents, and then it basically creates a really high-quality RAG system on all of your documents that you want or that you put into the librarian. So, I've done some pretty cool things with that. Like I took all of my Slack data. I've had, I think it was over 7,000 direct messages with people in slack yeah over the past six years and I've talked about everything from uh life love the pursuit of happiness yeah to hey you cannot post that in this channel I'm going to ban you so what is very cool is that i now can ask that librarian who do i know that's working on GPU optimization in Toronto I love it. If I've ever had a Slack conversation. And then recently I downloaded, there is a way that you can download all of your LinkedIn data. And so I did that.
Erica Hughberg (33:33) Who do I know in New York?
Demetrios (33:36) I did it all yeah who do I know in New York who do I know that's working on XYZ in New York and so you can ask the agent librarian that and it's much more than just the traditional rag because it can do things with the information so some things that it does is it will say okay well, I think these people are in New York, but some of these conversations were older. Let me confirm that they're in New York. So, then it goes and it looks on LinkedIn at their LinkedIn and it confirms that this person is in New York.
Erica Hughberg (34:13) Oh, amazing, like because I have this weird thing that I do in my WhatsApp like with all my friends I've now have like you know contact this and I grouped people like New York, London, you know. So, that if I go somewhere I can like look at my contacts like where do they belong in the world because being someone who's moved around in the world I think you also have this problem of not just like for work but even for like friends you forget who lives where and then you're really sad if you've been in London and then gone back home and you're like oh I forgot to catch up with this person but I actually really miss them but I forgot and so that's why I started doing that but the other part that it reminded me of like I had a guy who worked for me for a little bit years ago, and he used to have an Excel spreadsheet of every one that he knew. And he would have all of these like columns, right? To like explain like what they were working on, what they were good at. So, like he would know like, oh, if I need help with blah, I can go and reach out to this person. And I was like, that's so smart. Like in a way, I feel so smart because then he knew and he would be able to go back to people and be like, hi, I remember you told me you were working on blah, blah, blah. And I need help with this thing. Would you be able to help me? And it wouldn't be like completely random. Like he would actually have noted down.
Demetrios (35:42) Well, that's one of those things where you have to be very diligent about noting down every conversation that you have with someone yeah and now it's a lot easier because…
Erica Hughberg (35:52) Now I can do it so much
Demetrios (35:54) Yeah, it's it's one of these things where now we can between all of our recorded work calls and all of our interactions on slack or interactions on LinkedIn or Twitter dm’s or whatever it may be, you can kind of grab that information and then store it and have your own personal one, which is a glorified retrieval engine.
Erica Hughberg (36:22) Yeah. I think it's, but I do think what we talked about a lot today, which was, I think is so interesting around like how we use different agents, like that we didn't build, but we use them to be productive. And, I find it fascinating that maybe we won't see a massive boom in like agent, like people building agents. Okay, maybe it sounds weird, but I'm wondering if what we're going to see being built more over the next two years is integrations, where there's agents integrated with other agents. Or whether that is agents integrated with data, tools, governance, whatever it could be. Like, because a lot of these multi-purpose vendor agents can help you achieve a lot, right?
Demetrios (37:24) Yeah. Yeah. And you, we used to build agents, but then it became like we were building harnesses. Yeah. Less than building an agent, you would build a harness. And, I do think that some people still have a lot of dedication to the idea of we're going to have thousands of very highly specialized agents that can get called and can get used and you have this network of them because this agent just is does this one thing really really well and you can call it when you need that one thing done yeah I am not so sure about that world I feel like we are going to because it's me as the consumer it's going to be much easier if I can just use a generalized agent now maybe In the background, that generalized agent, all it's doing is calling these highly specialized agents. Who knows? But that would mean that we would have to spend a lot of time. crafting these specialized agents, right? And then you would have to have this boom of agent builders. And I think you make a great point here of, are we going to have a boom of agent builders or not? Because we thought we were. And then what we're seeing right now is that everybody's using agents. It's just not like we're building our own agents. We were integrating them with somebody else's. Yeah. We're using somebody else's agent at the end of the day. Yeah. And it's going pretty well. It is quite useful.
Erica Hughberg (39:09) Yeah. And I find it fascinating because I was just thinking about who builds agents. And I feel like it's very much people who are in specific, like specific companies that are driving that. And I'm not saying it's only the Frontier Labs. It's a lot of companies that do it. Like I mentioned earlier. But that's still not majority of companies. And I do have a funny thing that I found. I sent this article to you. It's from Gartner. It was released end of September.
Demetrios (39:49) What's Gartner saying in these days?
Erica Hughberg (39:52) Well, Gartner is saying that, Gartner predicts, I'll tell you the headline. Gartner predicts 70%, seven zero of enterprises will abandon a gentle AI built by vendor forward deployed engineering by 2028. So, it's an important, what does that mean? And at first, I was like, what does it mean? Right, because, so what. Why make a simple headline when you could make it really complicated, but wouldn't that make more people actually read it? Either they like, oh, I don't understand. Or they'd be like, well, I don't understand. So, now I need to understand. So they opened the article. I don't know.
Demetrios(40:34) It appeals to the intellectual.
Erica Hughberg (40:36) Yeah. But first of all, the funny part is that not selling like at this point in time, there's a lot of enterprises that, you know, get forward deployed engineering contracts with companies to come and help them build agents or agentic solutions. So, because it isn't that easy to get an agent that is built inside of your own little world into production. So, but what is a forward deployed engineer? It's just a rebranded consulting service, right? But it sounds cool. You're not using, you're not selling consulting hours. You're selling forward deployed engineers. So, first of all. Okay, I'm going to be silly, but maybe they will stop. 70% of enterprises won't use forward deployed engineers anymore. They'll switch over to consulting. No, just call it something else. That's not what the article says.
Demetrios (41:39) Let’s go back to the original terminology.
Erica Hughberg (41:42) It's like, yeah. No, that is not what the article says. But, they talk about this like. You can't rely on it. It feels so obvious, the entire article, in the sense that if you are going to have your own agents, you probably need to build them yourself and be able to own from creation. to like deployment and to outcomes as a company, like you can't have a third-party company doing that for you. And I'm like, okay, cool. Like, yes, if you have a consulting company coming to build you software, whether it's a Agentic or not, you kind of, you need to have an exit strategy so then you, your own company can own it. So, I thought that was, I feel like the entire takeaway from this article was. You need to view the forward deployed engineers as a delivery model. They help you deliver, but they do not help you own and operate. And I'm like, you did a whole study to conclude this? No.
Demetrios (42:45) It's like that meeting could have been an email. This study could have been a headline. Could have been one sentence there. Pretty obvious that everyone already knew.
Erica Hughberg (42:57) But I thought it was fascinating because you know how the heading, it was like, first I only, okay, so, earlier I saw this headline and I was a bit confused by it because I was like, we'll abandon Agentic AI built by vendor. That was like the part that like caught me at first. And I'm like, and then you continue forward deployed engineering. Aha. Okay. So now we're talking about something different. So, I feel like the headline is interesting. grabs attention because they're going to abandon something. And what does it mean? They're not going to use AI anymore. They're not going to have agents, but no, that's not what it meant at all. So, if you see this headline and you comp people there to open the article, they are not saying that they're not going to build agentic solutions. They're saying exit strategy.
Demetrios (43:43) I feel like I've seen this being touted for since. the dawn of ChatGPT where folks are like we need to have our own models we need to use the open source ones instead of using oh that's another frontier and yeah and this in a way is that it's like saying we need to own our agents we need to use some kind of open source model because you know that they're not going to go out there and train a large length model and so you use some open source model you use it on your own hardware yeah whether that's on-prem or you have some Neocloud and then you're building the agent harness or you're grabbing some open source harness and you're building this agent from the ground up that just it's like yeah that is great in the world uh in a utopian world but as a friend of mine Sahar was just saying when I was talking to him and he was like man I'm pretty scared considering all of the risks that are now out there with this cyber security stuff because I'm just pushing so many things into production and not really thinking twice about it but and then I said yeah but aren't you doing the adversarial reviews with a different model and he's like of course I know I should do that. But, are you doing that? Because I'm not doing that. And no, you know, logically, theoretically, I know that I should be doing that. But to actually do it and set up that loop and constantly have some adversarial PR loop going on. It's a lot of work to put into a loop. And then he was like, I skip it.
Erica Hughberg (45:40) I, I tried really hard to have like this. Okay. So where, where have I landed in my own workflow sometimes with this sort of process of having an adversarial, like a reviewer. The problem with code is they can always be better and it can always, and I felt I got stuck in that it was never good enough. The funny part is that if a human had done it and a human had reviewed it, it would have been good enough. Like, and. I ended up like being like no I can't release any code it's all bad like nothing's good enough like that was like a few weeks ago where I felt really stuck because I had a review after review after review, review, fix review fix and I'm like, but, when does it end? When are we done? When is it good enough? And then the problem is that there's a level of coding that's a piece of art, right? How do you decide when it's good enough for the job it's designed to do? You need to take informed risks. I feel like all software is a... From a security standpoint, it's about taking informed risks, right? Whether that is from security or stability, like sometimes you're taking a stability risk because you feel fairly certain you're not going to have a ton of users and you have some other middleware that is going to throttle and like, you know, block traffic, right? You are taking some calculated risks that, yes, what would a DDoS attack look like? Okay, because we have this middle layer that reduces the risk of a DDoS attack. So, therefore, we can take in calculated risk on this other lower level part of the stack. But I feel in the world of AI coding, it often sees the code, not the system. Therefore, the same informed risks we may take as humans get lost.
Demetrios (47:51) That is a great point. Because if it's just hyper focused on the code, it is not going to think about these trade-offs that you, as a human are thinking about and it's going to constantly try and optimize the code and as you said there's never a point where it says oh this is good enough let's stop and go on to the next thing.
+ Read More
Watch More
Building AI So Everyone Can Flourish
Posted Oct 26, 2026 | Views 214
# Tim O’Reilly
# O’Reilly Media
# Agentic AI
# AI Agents
AI Agents Are Revolutionizing E-Commerce at OLX // Nishi and Beatriz
Posted Nov 22, 2024 | Views 1.2K
# Olx
# Prosus
# AI Agents
# Agentic
# GenAi
Before Building AI Agents Watch This (Deep Agent Expertise)
Posted Sep 05, 2025 | Views 689
# Context Engineering
# Prosus Group