1. Podcasts
  2. Agentic Conversations (formally mlops.community)
  3. From Slop Code to Software Factories
Agentic Conversations (formally mlops.community)

29 Oct 2026 · 47m · 20 chapters

Episode artwork

From Slop Code to Software Factories

In short

Two hosts ride a tandem bike through Amsterdam while discussing “software factories,” why current AI coding models still produce “slop” that breaks at scale, and how agent tooling (computer use, MCPs, browser agents) changes software development and workflows.

Guests

Dex (the main guest). Background: builds “software factories” at Humaneer; works with large-company engineering teams (often 100+ engineers, sometimes 300+). He focuses on agent/IDE infrastructure, real-time sync, tool reliability, and review/approval workflows.

Key claims

  • Models diagnose bugs better than they design maintainable abstractions; “fixes” still turn into spaghetti.
  • Browser agents and MCP/API tools are the “gateway to the machine,” but speed and profitability tradeoffs remain.
  • Tool reliability improves when tool schemas match what frontier models were RL-trained to use.
  • The industry needs reusable “wiring” building blocks (CI/agent orchestration), not everyone reinventing it.

Notable examples

  • Computer-use automations: flight check-in boarding passes; government license appointment booking.
  • Demo pipeline: generating storyboard/contact sheets from raw footage using Opus 4.8.
  • Daily briefing workflow: agents gather 100 items, then narrow to 5 for a newsletter with human “why it matters” takes.

Written by AI. May contain mistakes.

Chapters

0:51

The Joke Setup

Hosts attempt to talk about software factories amidst cycling challenges.

“Way harder than I thought it was gonna be.”
2:14

The Coding Slop and AI Models

Discussion on messy code and the expectations from AI models in coding.

“Let's start over because I completely lost the train of thought on the joke.”
4:25

The State of Software Development

Exploring the improvements in AI for coding versus real software development needs.

“You realize that you being in the back, you have to do like 90 % of the work.”
6:42

Utilizing Computer Use for Efficiency

Discussion on using AI tools for managing tasks and emails effectively.

“What's like a cool new thing you figured out?”
8:34

Real-World Applications of AI Tools

Hosts share practical experiences with AI tools for personal organization and tasks.

“I think it's also, yeah, when you start and stop pedaling, it's like if you start slow and end slow, it helps us.”
11:47

Challenges in Developing Good Software Tools

Discussion on the challenges of creating effective software tools and MCPS.

14:00

Exploring Tool Recommendations in AI

Learn about the importance of tool recommendations in AI and coding.

“Like for cloud code it's a lot of the coding tools and so it's like I don't know we built several harnesses internally we like keep like building them and then like throwing them out and rebuilding them.”
15:00

Building and Iterating on Coding Tools

Discover the iterative process of developing and refining coding tools.

“We'll try and grab this spot in front over there maybe.”
17:30

Generating Video Content with AI

Learn how AI can assist in creating video content from raw footage.

“Like if I'm doing like deep, deep architecture coding, or even on the front end doing stuff that is like data plumbing and stuff.”
21:00

The Role of Agents in Software Development

Discuss the implications of using agents in software development workflows.

“I'm like, cool, here's, let's make these changes.”
Show all 20 chapters
24:20

The Future of Browser Agents and MCPs

Explore whether browser agents will replace traditional software in the future.

“whatever it may be, that's kind of like implementation detail that I don't necessarily want to have to care about.”
28:00

The Shift to Headless Systems

Discussing the transition to headless systems and its implications for product development.

“reason to have a human-facing site, then why not just make it as easy as possible?”
29:14

Automating with Codex

Exploring how Codex can streamline workflows by handling tasks like form filling and email responses.

“One is just booking out things like, hey let me book some museum visit.”
30:56

Enhancing User Interface in AI Tools

Discussing the need for better editing capabilities and user interfaces in AI applications.

32:36

Improving Communication with AI

Analyzing how AI can enhance communication and reduce information overload during teamwork.

34:28

Creating Effective Newsletters

Describing the process of curating and writing newsletters using AI assistance.

“This was the skill we did, the open source skill we just published.”
36:36

Building Software Factories

Investigating how companies are adopting software factories to optimize engineering workflows.

“Now, the end state is I want to get them into a newsletter with a little blurb, and then I want to comment on why I find this important.”
40:06

Challenges in Software Development

Discussing the challenges and solutions in creating efficient software development environments.

42:03

Building a Software Factory

Learn about the essential components and challenges of creating a software factory.

“And you, for you, it feels real time and you can run that demon anywhere.”
44:35

Overhead and Efficiency in CI/CD

Understand the complexities and overhead associated with CI/CD systems and how to streamline them.

“There's like, it's like the overhead of building a CI.”
A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On Agentic Conversations (formally mlops.community), what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00Thought it was a brilliant idea. Oh boy. Oh boy. Oh boy. Oh boy. It sounded amazing when we were planning this.

0:09Dexter Horthy: Keep going. You power. Do a tandem podcast, they said. It'll be fun, they said. Oh my god. This is way harder than I thought it was going to be. Dex is huge. He's going to look like a freaking giant on the back of this tandem. Dude, that was... Okay. Well, here we go. Stoop waffles, canals, cheese, tall people, space cakes. What is more Amsterdam than smoking weed, riding your bike along the canals? Well, today we're doing that with Dex. The only catch is we have one bike for the two of us. He's going to tell us all about how he is building software factories at Humaneer. Let's do it. Beautiful.

0:52Beautiful.

0:57One, two, three.

1:03Whoa, why are you going that way? Way harder than I thought it was gonna be. I'll see you on the other side. You're good at it, man.

1:10Dexter Horthy: I don't do anything, man. So thinking that we were gonna be able to talk about software factories while doing this. Bro, we're gonna have to just talk about fucking staying alive out here we're gonna let davies be the lead support we're just following you dude all right we're not oh yeah there we go yeah now we're going okay so now tell me all about astra there was a joke we used to make uh when we were coding with like clod 3 7 and opus 4 and some people would kind of be like yeah, man, I know the code is slop but like By the time it matters. This is very mentally intense right now. Oh Okay, one two three go back on all right back on this feels a lot like the first time I rode a regular bike I got a lot you got to just like believe yeah All right, hopefully it's just chill roads from now on and you can tell me about this joke.

2:14Let's start over because I completely lost the train of thought on the joke.

2:19Dexter Horthy: I'm glad you picked up that it's a joke. I haven't even gotten half the setup. Give it to me again. So, we used to say this thing when we coded with like Claude, Sonnet 3-7 and Opus 4 back in the day of like yeah the code's kind of messy but like I'll just not read it yeah and that'll be GPT sevens problem the the impression being like there will someday come a model that's good enough that is so good at coding that'll be able to fix all this bullshit yeah and I should just focus on vibing and getting it right and figuring it out etc yeah as we go doing coast yeah GPT 3.6 just came out and uh or sorry gpt-6 six yeah gpt-6 astro just came out it's not that much better at coding i saw a couple people i saw which is weird because they're leaning into that like you gotta know that that's like one of the main use cases i mean it got better this actually i think this is a take from uh martin cassado over at andresen he said basically like it got better at some things it used to be bad at like on the right we're going it got better at some things that used to be bad at like blender and uh what is it like uh doing game slop yeah like asher can make video games and it's like much better computer use and stuff like that is actually pretty brutal that's true i have seen that somebody created a whole uh presentation but instead of creating a powerpoint It just said create a game engine that could then be used for a presentation And so the game was walking through Like this whole thing like a room that got the point across.

4:10Yeah, so We are here now and it hasn't gotten better. So it's still GPT 7's problem

4:16Dexter Horthy: Well, so that's why I take it like we're already at six and from five to six like yeah We got better at things, but we didn't get better at building maintainable software We got better at like diagnosing bugs but the fixes to those bugs still end up being quite uh we're clear we're clear we're clear let's do it oh boy you gotta lead bro you gotta lead because I don't really know where I'm going coast it coast it coast it all right so anyways that's that's the take is like yeah they're better at like breaking into systems and diagnosing problems but they I haven't learned how to build good abstractions or design code bases in a way that doesn't turn into spaghetti.

4:58I think we're going left here. Left here? Let's do it.

5:01Dexter Horthy: Power up the base cannon. Yeah, there we go. You realize that you being in the back, you have to do like 90 % of the work. That's okay. We're going to stay straight with these guys. I think we want to go right. well straight that was a close call everybody in Amsterdam fucking hates us today all right you ready one two three oh fuck we're starting this one one way you got it yeah we don't we're going left oh we're going left all right hard left hook I love how it's nothing for most people they're just like fucking tourists yep and we're like struggling to stay on this damn bike for us it's life or death over here meanwhile some guy's got a couch on his bicycle yeah it's like i drop it off a bunch of like speaker equipment he just had this giant rack on the front like doing av stuff that um fable 5.1 cannot do yet that is for sure so other things fable 5.1 cannot do yeah or maintain a code base as you have pointed out and we're just continuing the slop but my big question is if it runs does it matter i mean i've said this one this this has been my drum i've been beating for six months now but like it may run today and it may run in three months and then you have hundreds or thousands of users or millions of users and then one day you wake up and you're like I haven't read the code in three months and why is something is broken and you shout at the bottle and you just keep pulling the fucking slot machine and it just can't fix it you don't know why you're sitting there and you're like what is oh I'm gonna start oh yeah sorry I forgot we have to be in sync yeah it's like a fixie all right we're good We're chilling.

7:10Dexter Horthy: We can keep coasting. Yeah, we'll coast some more. All right, here we go.

7:21Dexter Horthy: I don't know, man. What's your coding stack these days? What's like a cool new thing you figured out? Dude, well, so you know what I am doing a lot is ever since Jason Louie from OpenAI came on the podcast and told me about computer use, I'm all in on it. You just like, Codex, go check my emails for me and stuff. Dude, I've set up so many beats that it will do. And one of them being, here, hold on, we got to stop. One of them is like, check in on flights for me. Yep. So, oh, oh. Got it. So this morning I flew over here. It's left. You go left? You went that way. Yeah. Why? You told us to turn left.

8:04Why would he say that?

8:06Dexter Horthy: All right, let's go catch him. All right. All right, one, two, three. Oh, boy. Sorry. Yeah, you got it. You got it. You got it. And we've got some water. Let's do it. Water feature. Water feature. So Jason Louie came on the podcast, and he told me about how good computer use is. Yeah. And since then, I've just been setting up all these scheduled tasks or recurring tasks. like going through all of my different emails and so I don't have a problem with oh there's one connector that I can use and then I can't connect the other five emails I have or going through all my slack workspaces and it's the same thing I can have it go through all that stuff and then the one that's beautiful is when it needs to go and do stuff the fucking ability for it to fill out forms yeah is unmatched so I don't have to worry about oh is it able to get that information from an API not and you can just go and do it as if it's just impersonating yeah yeah because it's already I'm already like kind of signed in on the browser I think we go right here so right see if we can yeah let's see if we can he says right right right oh boy well we'll go on the road We'll go in and it will cut in.

9:31Dexter Horthy: We're getting the hang of it, man. It just took a minute. I think it's also, yeah, when you start and stop pedaling, it's like if you start slow and end slow, it helps us. It makes it so much harder because you're going one way and I'm going another way. Is this guy going right? Yeah, yeah, yeah. We're clear, we're clear, we're clear. And so one thing that I did that this is full Jason telling me he set it up and so i set it up and it's just so easy to set up anytime you see it so i have like a recurring task that goes through my email and anytime it sees an email from a flight company like a lufthansa it just goes and it grabs my boarding pass so if it says like your check-in is ready it'll go and it'll fill out the form and it'll grab my boarding pass and so now this morning when I came here it was beautiful because I forgot to check in as I do and when I was looking for the email I didn't find the email that said hey it's time to check in I found my actual like boarding pass in my email nice so those are the moments where you're like damn that is just making my life so much more simple and I did the same I've said this again with like I did it with bunch of german legal stuff so me having to set up appointments with the government yeah i just have it go and do that like for your llc or whatever yeah or just like uh i had to get a sorry i had to get a um what did i have to get yeah we can coast we can cruise just cruise it yeah i had to get my license renewed and so i just had it set up set me up an appointment to get my license which was beautiful because all I got was hey your appointment is on this day here's your number and like bring these documents and then it created a calendar invite watch out and I was good so the cool thing I had I was actually out of fable credits yeah and I wanted to do it with the Claude model all right so I ended up using opus 4.8 which means I can't wait to see what happens when we throw aster at it or whatever uh but i've like i've been procrastinating making this demo of human layer yeah and it's kind of hard to demo because all the cool shit we do is like very collaborative it's like so like to do a good demo i need to be in the middle of demoing and then be like hey kyle come comment on this thing or like come like co-prop this session with me so it's like it's like a two-person demo and so i recorded all the footage i needed but it's like to edit it watch out on the left all right hold on we got a right too so yeah there's like a ton of downtime uh and i'm like i don't want to go read through all this stuff and like throw it in davinci or whatever and like and then have to try to put it back together so i hooked up uh there's this new editor called palmier uh-huh uh it's some yc company but it's basically like davinci resolve but with an mcp beautiful which davinci just set up their mcp do they have an mcp now yeah which i'm very excited to try and i whenever i see non-tech company or like non startup-y tech companies try to make mcps like the first three versions of them are always really bad yeah it's like oh you need someone who's like super ai pilled and has used a thousand mcps and knows exactly what makes a good one otherwise i'm do you kind of have to have like eval your MCPs don't you that is a thing that because there's a lot of MCPs out there there's not a lot of good MCP servers yeah it's I mean because you just like to design a good it's like you know Coda Hale has this bit it's like just like an axe must fit the human hand to be useful software must fit the human mind to be useful and like not a human mind is it exactly so there's like another take of this which is like mcps must fit the agent mind to be useful and like if you don't know how agents think and like what kind of tools they're good at calling and not good at calling then it's unlikely you're going to nail the mcp on the first try yeah that's one thing that i've seen folks doing is the whole like oh how can we make the tool calling more reliable on these mcp servers and so you have this stuff like tool search but there's almost like this next layer that you need which is tool recommending because you're like oh hey i need to have this this person match with these tools they always call they really like this type of a thing but anyway i'm getting off track what did you do with well and just one more thought on that it's like i also think there's a lot to be said for like months and months and millions of dollars of compute being spent to RL the model on a specific tool set.

14:39Yeah.

14:39Dexter Horthy: Like for cloud code it's a lot of the coding tools and so it's like I don't know we built several harnesses internally we like keep like building them and then like throwing them out and rebuilding them. Really? As we like learn about the problem space. But every the thing we do in all of them is like if it's a cloud model we make all the tools that are like mirror things in cloud code look exactly the same as if it's running in cloud code so that you don't have any issues like well you just take advantage of the you know again like millions of dollars and months of RL training that the labs are doing to teach these models to be really really good I turn it in a certain way yeah and like codex is there the codex edit tools like this patch thing that does everything and if you try to give codex the like old string new string patch tool that you get in cloud code it's just way worse at using it and so it's like how can you design your tools to look like what the model is RL to use and the way you do that is you look at like okay in the frontier lab like flagship harness how are the tools exposed to the model then we can have that reliability and we can know like is it going to work well there's a higher probability chance it will because yeah it just fits in with the schema that we you know has been yeah the model just is like no it wants to call things in that way it's been we can cross and then go left okay we'll cross and go left and the thing that i was i think we're clear oh rarely is it this easy uh but boom nailed it well done so dude i love this town i freaking love amsterdam and when it's sunny and you're on a bike it's even better oh here it is we can stop here throw the bike here in this parking spot beautifully done we made it dude we survived was a little bit more intense than i thought it was gonna be i wasn't expecting so much out of it.

16:41We'll try and grab this spot in front over there maybe.

16:45Dexter Horthy: Or just... We can get two chairs. We can figure it out. Should be simple. I saw something on Twitter which was like, the post was like, he'll climb this until it's professional quality. It was like a prompt someone gave to something. And so I gave it all this raw footage of just me using it. It was like, not hours, but at least two hours of footage to make a one-minute or two-minute demo. and so I was like here's the outline of the demo I want to make here's a bunch of footage in a folder like go make contact sheets of it of like you know every two seconds like so you can see in one big in one image like token efficiently you can see what happens in every clip and then zoom in on the parts that you care about and then here's an 11 labs key and like basically make me a story voiceover so I was like I'm gonna do the final voiceover but I'm like I want to hear how you think this goes and like help you like time it like and pace it exactly to like where things are happening and anyway with 4.8 this is an opus 4.8 not even like a super dope model right I mean opus 4.8 is pretty damn good before up until opus 4.8 I was a 4.5 guy I didn't really you didn't go to the 4.6 I mean I tried it a little bit I tried 4.7 was like super spazzy 4.8 was like It was good at the kind of things that I use, Claude.

18:04Dexter Horthy: Like if I'm doing like deep, deep architecture coding, or even on the front end doing stuff that is like data plumbing and stuff. Turn around. Look behind you.

18:20Dexter Horthy: Oh, that's funny because he could hear us.

18:25Dexter Horthy: Data plumbing stuff, I will always just use codecs pretty much entirely. Isn't it funny how we have our likes and dislikes for the different models? Like, oh, if we're doing data plumbing, we'll go to one model. If we're doing... Whenever I'm doing any kind of brainstorming, I'm always flawed with... And usually it's fable, if I can, with the superpowers skill. Okay. And just because I've come to like that one the most, who knows if it's actually the bet? Yeah. right but it's less about what is perfect and more like what do you have good intuition for and like what do you it takes time to learn how to use these things yeah yeah and some of them you just get so frustrated with because you have to as you said like steer it in ways that why are you doing it that way and maybe it is actually okay to go that way it's just not how you would necessarily do it and so you're out of your depth quicker and then you're like Actually, I like this other model because it does it closer to the way that I would want it to do.

19:30Dexter Horthy: Yeah. I usually tell people, like, don't waste time, like, min-maxing across Clawed and Codex and Cursor. Like, don't try to, like, it's really, build a really deep intuition on one model. And then maybe you can start bringing in other ones. But if you're just constantly switching, you're never going to kind of get to know any single one. And you're never going to, like, know its strengths or be able to really depend on it. and have a good idea of what's gonna happen when you say a thing. But anyways, yeah, so we're using Opus 4.8, it was like make the contact sheets, then make me a storyboard.

20:01Dexter Horthy: It was like HTML doc with two or three screenshots and the exact words that go on the talk track. And then two or three screenshots, and for each beat, and there's 10 beats in this one minute video. Maybe like six. And then I was like, cool, I did a little bit of iteration on the storyboard, and then I'm like, I'm going out for a run for an hour. Like, and I just put it in a loop of like, he'll climb this until it's professional quality. Compact. He'll climb it. I just like queued up like 10 messages for it to just keep working while I was out for an hour. I came back and like the demo was like low key.

20:33Dexter Horthy: Like it sucks. Like it's, it's, it's very slob. But like I was impressed with what I could do. It's like, it's not exactly what I would have shipped. But I was like, this is like good enough that I'm just going to start recording my screen while I'm working. Knowing that like. You can find jams. Yeah, and I gave it, I also gave it a like speech to text key so it can see what I'm talking about while it's on screen. And so it's like, yeah, it's slow, it's inefficient. I've never like sit there and work back and forth. But now I kind of have my like, you know, AFK workflow of like, here's all the stuff.

21:04Dexter Horthy: Go make me a storyboard. And then I go away and do other stuff. And then I come back to the storyboard. I'm like, cool, here's, let's make these changes. Cool, go make the video. And then we like iterate from there. it makes me think that it could pull out the gems and it could give you insights that you don't necessarily realize are there sometimes where it's like oh hey this is happening and you didn't even realize hey that's a novel thing wow okay cool I should talk about that more yeah the next is you're so in the weeds that you don't see the force from the trees you know yeah yeah being able to take like I mean people they're like extract cool clips from a podcast AI has kind of been around for a while which I've never found one that is so good that I'm like wow I gotta pay for this yeah I tried a lot of them and it's not quite there yet but well but now I can run this from my Claude which means I can own all the preferences I can own the Claude MD I can write the skills myself which has always been the like I know a lot of people are saying it's like I don't want to use your agent I want to use my agent and like point at your MCPs don't ship me a like mediocre agent when I have the world's best harness at my fingertips yeah and my preference right like what I like but then again it goes back to the whole thing with if I'm using your MCP and going to be able to do that is your MCP any good that's to be good yeah the MCPs have to be good but it's like instead of investing time and like learning how to do sandboxes and running agents and like do multi-tenant agent cloud or whatever the hell it is it's like just build really really good tools for my agent to use well and that's one thing that i've been thinking about especially when it comes to browser use and how if you think about how we used to build software and create software it was very much because somebody was not happy with how things were being done and they would think man this is a pain like i think it could be better yeah and then they would build that world that they envisioned okay but now if agents especially browser agents are doing it you don't have that pain yeah it's in a different place it's as long as the browser agent doesn't get lost and it's like constantly texting me like hey I need to go log into this page or whatever it's okay yeah so then you're not having you can have shitty software out there but it's not going to be like oh we need to reinvent this software in a way that is going to revolutionize things because if the browsing agent, again, if it does it, then why do you need new software?

23:41Dexter Horthy: I mean, okay, so here's my question. Do you think, like, browser agents are a band-aid? Like, do you think the future is everyone's just going to have MCPs and agent tools and browser agents is just, like, you know, for the next two or five or ten years that it takes us to get there? It's going to take a while. It's going to take a long time, but do you think we'll always have websites? or do you think in the future I'm just going to talk to Claude or Codex or whatever? It's the gateway to the machine, of course, whether it's local or like you're just going to talk to the chat box. And whether or not it is doing things through the browser or it's doing things through the MCPs or the APIs, whatever it may be, that's kind of like implementation detail that I don't necessarily want to have to care about.

24:25It's just going to happen, right? Like, oh, it tries first with an MCP, it's not able to do it. Cool. Like, is there an API? Is there a public API? Do I need to pay for the API? Because that's another thing with browser use. Like, you don't have to pay for it anymore in a way.

24:41Dexter Horthy: But you also can't get the speed that you used to with APIs. Yeah, you can't get the speed. And also, like, will all those websites actually still be profitable if, like, they were making money off of ads and the agencies, the ads, and doesn't care? that or think about and I've been talking to a lot of folks about this if my agent now doesn't care about you up selling me on the e-commerce store yeah and all of this stuff that we've built for humans over the last 20 years to fully get the most out of their session yeah now the agents they just blow right by that I mean my agent has tried to upsell me before it's been like oh I was on this site like shopping you know I was testing a workflow of like cool send me the pictures of this thing I'm shopping for and then I picked the one I liked and it's like hey by the way if you buy like two more things you get free shipping do you want to do that and I was just like did you the same answer I always get which is no but like but the agent did like the upsell did transfer from from the agent from from the website through the agent to me like I did I got the pitch again which I was like yeah that's probably still gonna be a thing yeah that's true or there's gonna be different sneaky ways that it can come through with the agent but I mean if you said to the agent no and don't ever do that beautiful

26:02Dexter Horthy: thank you this is my favorite cafe dude I love a good pour over so the other thing though because I I really like think about this a lot with the browser use for the computer use in general yeah Do we have it now and then later everything's MCP and API? Or is it always? Because I just don't see the German government in my little small ass town adopting MCP anytime soon. So to fill out forms and stuff. Okay, maybe it's in 10 years. Like you said, maybe it's like they don't do it now. just like not everyone had a Facebook page back in the day but then everybody had to go and get one when Facebook was the one spot that you had to be.

26:51Dexter Horthy: Well in the States in like 20 it was like 2012 or something it was a really big deal when we got healthcare.gov it was like oh you can now go get your government benefits on the internet you don't have to go fill out a form you don't have to send in a letter it was like a huge deal so probably have something like that which is like you know gov MCP or whatever it is but again I don't know when it's gonna come right like it was in the in the US it was you know 12 years after the like World Wide Web became like oh this is the future of the economy and information and communication but my big question is would that necessity be there to create the MCP if browser use can already do it that's where I'm I'm going to I know you're going to like hey everything's moving towards MCPs yeah because it's obviously the better way of doing it but But if you don't have that catalyst, because it can be done already.

27:45Dexter Horthy: Oh, it's beautiful. Thank you. Oh, nice. That's okay. We're very thirsty. We'll figure out how to drink it all. Yeah, I mean, the question is like, because today websites are great because they serve a dual purpose. It's for humans and agents can now use them reasonably well versus like, if there is no reason to have a human-facing site, then why not just make it as easy as possible? Yeah, everything's headless. But that is easy for things built today. But I feel like anything that is, A, doesn't have a huge tech team behind it. Yeah. And B, doesn't really see the need in it. And they're not fully convinced anyway.

28:27When are we going to get to that point where those folks switch over to fully headless? Just your product is your MCP server, which is wild to think. your product is not anything more than how your MCP server can then leverage like MCP apps to create a better experience inside of my yeah yeah I mean

28:48Dexter Horthy: you basically like become some you either become like an MCP over a system of record or an MCP that is interface to something in the real world yeah or it's like moving physical goods to my location or something like that or like booking me a hotel or whatever it is. I've tried really hard to have the experiences so far on doing everything I can with browser use, right? Yeah. One is just booking out things like, hey let me book some museum visit. I was just in Tokyo. Yeah. I booked a bunch of museum visits and that was fine. Like anything with it sounds great i literally my mind switched when jason said i never fill out forms anymore codex go do this for me from then on it was oh my god if i'm ever finding myself filling out a form i'm doing it wrong i need to figure out ways to make it where my chat is my command center and one thing that I did that a few friends thought was kind of novel so I bet you've been doing this for ages is you get your daily briefing which is not that novel whatever it checks all of your emails all your slacks all your I have it check whatsapp and telegram and signal and then I also have it check YouTube to see how we're performing there yeah it'll give me and it can do that with browser use again like it can look at your YouTube analytics yep and check all of that and so the way that I respond to it is I don't then go oh Dex sent me this email I should go and respond in Gmail I just turn on whisper flow and then I say okay respond to Dex and say blah blah blah yeah this is one of the best workflows that I found whether it's like reading a plan of what an agent wants to write in terms of code or whether it's like hey look like here's a giant report of stuff you just turn on whisper floats think out aloud as you're reading through it and then like send it to the model and boom and it comes back if you need to iterate a few times you can one other thing that I've been really enjoying and I have been talking to a lot of people about this again the codex app yeah yeah the desktop app sometimes it will give you back a response that is a card that you can edit in app like in the chat so I've seen that the last time I tried it was super buggy so I tried to edit it and it sent the email I was like no dude there are 80 engineers on this project just on the desktop I have a buddy who works on codex desktop now so I send them all my bug reports good as you should and you should so I think they fixed that because but what I don't understand is why they don't have it come up more yeah and the editability of the chat or the responses from the models is a nice ui pattern that i want to see more and more you don't you're you're you would rather just edit it in line than just be like change it to be like this well sometimes it's it's a few words that i want to change yeah so you don't want to be like copy the thing paste it in and then paste the other thing you want to change it easier just to say like delete that whole sentence there by deleting it as opposed to yeah doing that workflow and so that got me thinking well why don't we have document editing inside of our chats and why don't we have like you just showed me more commenting abilities inside of our chats why don't we do what about notion drag and dropping text blocks yeah or notion like being able to call a database in line yeah this is this is the next feature I want from my like silly video production pipeline is I want Palmira doesn't have this today but I want to be able to comment on the timeline like whatever there's always some thing that you're working on whether it's the code or a plan or an email like you want a really high bandwidth way to just be like this time stamp here's my comment without having to say at this time you did this and here's what was happening and here's how I want it it's just like make this frame.io does that exactly yeah yeah it's for other humans yeah you want it so that it's an agent that can just know hey right here yeah take that part out I mean what I'll probably end up doing if Palmira doesn't ship comments on the timeline is I will probably be like cool now upload it to frame like the workflow is like okay when it's done do an export and upload it to frame and I'll comment on that and then pull the frame comments like I think I think like the I mean this was kind of like early days of human layer like the super broad idea was like how do we give really high bandwidth high quality like review interface and part of it was like the model should text me when it needs help but also like I should be able to click a link and interact whether it's like data labeling and I want to like approve or deny it or whether it's a classification or it's an email to send or whatever it is like all those things kind of benefit from like purpose-built UI is for the task at hand yeah yeah exactly and that is one other piece that I've been thinking about which is context is that context is just as much about oh this means this in this organization yeah and it is also when I present this thing to Dex I present it with this UI yeah because it makes his job much easier it gets him from point A to point B faster.

34:28Yeah. That is also context.

34:30Dexter Horthy: Yeah. And like the default of most agents, I mean, this is why I like, I struggle with chat and I think we need something richer is like people have been complaining for months about like Fable just info dumps an entire wall of garbage and you're like, give me something better. This was the skill we did, the open source skill we just published. Yeah, show me. Yeah, show me. It's just like, I ain't reading all that. Like, make it more visual and it's like. Or bro. Yeah. It's the same thing. It's just like, come on, man. What are the, what are the things we can do to maximize the bandwidth of human, both in terms of the volume of information that can be transferred, but also how do we make it use the parts of your brain that don't require a lot of energy, right?

35:10Dexter Horthy: If I ask you, this is an old Code of Hale talk that I recommend to everybody, but it's just like, if I ask you what's three plus six, you don't decide to add three and six. The number nine jumps into your head. But if I ask you to multiply two three-digit numbers, like your pupils dilate and your adrenaline flows and you're like let me focus on this shit like yeah if you try to do in your head and then it's like and also like you're exhausted at the end and so it's like how can we how can we how can we make things like how can we optimize the way our tools communicate to humans in a way that like uses the parts of your brain that don't require effort do you feel like that is as much of understanding the whole end to end or maybe not even end-to-end just the next step and then letting the model optimize for how it presents it or you it's a back and forth between how the model presents it and how you say you want it you're talking about like well I'll give you an example so we do the daily agenda newsletter and the way that we create that newsletter is we have ten different agents scouring all the different spots whether it's Reddit or X or just top news sources.

36:23And it will bring us back 100 different things that are happening today. So it's like a personalized news feed. What you could do is say, okay, well, I've got these 100 different pieces. Now, the end state is I want to get them into a newsletter with a little blurb, and then I want to comment on why I find this important. Yep.

36:48Dexter Horthy: so are you going to do all 100 or do you also have to kind of narrow it down yeah exactly just five yeah okay we do five every day and so you take those and you say these five are the ones that i want in there then again this is why i come back to the context is also can i create a gui that will make it as easy as possible for me to just say i want this this and this and say that these things here give me the one or two sentence blurb and then yeah my burb is is this part so you've got like the overview blurb and then you've got why i feel like it's important the human take and then you want to be able to tag that up and make changes either edit it directly or put comments on it yeah so the first part the first step is i need to take all of these hundred and i need to whittle it down to five that i like and then from those five they need to go into the newsletter sending software yep it takes a few iterations but you know like all right here's the whole workflow potentially there's a world where I'm copy and pasting this one and then I'm adding the link to it and then I'm putting it into the newsletter writing software and that's really cumbersome yeah or you build out that context again you connect these two sources where I've got all of the different agents feedback that or the articles I should say I've got all the articles and I can just say select and then boom it gets piped in and it's in my my newsletter software already and then I can just go and it also has the two sentence blurb and it's hyperlinked yeah UTM parameters or whatever like all that stuff is pretty basic table stakes for a newsletter sending software and then I just go in and I add my take at the bottom so what I'm trying to say is that whole end-to-end type of workflow yeah I had to get better and better at but I also have to know how to present it to me like you're saying so that three plus six is nine and I just go boom I like that oh that's cool that's novel boom that one's in there you don't have to like squint your eyes and be like what is this I don't know what it is how do I figure it out let me copy and paste it into here and let me try and like think of something novel to say because i'm really trying to understand it yeah so there's that but then again what i keep coming back to is does that even matter if we get good enough at the chat being the interface yeah i'm not going to have a software that sends emails you're not going to have a software where i coagulate all this data I'm just going to have the chat and chats gonna have little mini apps in there and some of them are just created on demand by the agent and so that that's kind of where we are right now is we have these softwares that we would put as stop gaps yep and if you do something enough times it's repetitive enough you have this workflow that you're comfortable with and you can optimize the workflow yep and plug in the human where it needs plug in however how long will we need that for right exactly how much of this stuff and it's like okay cool there's always this tension between like how do i solve problems today versus like what is worth building like how much effort am i willing to put in if i think there's a 30 chance that this will be not relevant in a year or two years or 10 chance in four years or whatever whatever the like probability curve is well how do you because human layer is very much in the bloodbath space of building yeah everybody's building in the coding space because it is the up until now the most lucrative well and it's the it's the one with the best like like just like problem like solution fit basically it's like it actually uh it works well uh i think actually like what we've learned a lot of in the last couple months working with like we work mostly with larger companies like our like ballpark is like at least a hundred engineers but most of the teams we work with have like more than 300 a handful of seed stage companies that love it whatever they're just like you know fans but what we're finding more and more of is like every every principal engineer inside every company is kind of going on the journey of like building their own personal software factory and then like a lot and then there's like well the CTO wants background agents and they want the like ramp inspect or like whatever all these like companies are bragging about this stuff everybody wants to be able to do that ramp ruined so many fucking software yeah that damn X post that they put out yeah everyone's like oh my god we need to be like ramp we need to be like Brex we need to be like work OS whatever it is and so what we actually found is like having a full stack IDE is like it's very useful as like a starting point and it's like it's nice in terms of a bridge of like cool people are used to sitting on their computer and talking to an agent and writing code or writing coded an IDE or whatever it is and then everything after that is the same that when it's done you make the pull request and like maybe there's a gentic review over there but like basically the interface between the developer and the rest of their team and the production system is kind of the same but then what actually I think what what what the industry needs is like these more like there's parts of building your own software factory that you should absolutely own especially like defining your dev environment and how the code runs and things like this but there's parts of it that are very much I would call like undifferentiated overhead of like cool I have an agent running in a container somewhere sound sandbox or whatever it is we picked a stack for this and we can we have it running how do I stream messages in and out of that right like how do I hook that up to slack or a custom UI or whatever it is show what the agent is doing let the user do approvals and send messages in how do I how do I like when the agent is editing code how do I take that diff and make it available to the team so that I don't have to wait for pull request it's small things and it's like it's very easy to vibe it out but doing it in like a scalable way that's super real time.

43:06Dexter Horthy: It like actually ends up being, this is like, it's not even like, you don't even see it in our product, but we spend so much time building this sync engine so that like when you're running it locally, you're actually like, your UI is talking to a cloud, is talking to a demon running on your laptop. And you, for you, it feels real time and you can run that demon anywhere. And so like, we spend so much time making that sync experience. We're building on the electric sequel and durable streams and all this kind of like next-gen real-time streaming tech and it's surprising how much that like it's just a little bit unwieldy but doing it in a very effective way where you deal like client registration along polling all those like stuff that has been like kind of solved but operating it is hard and so it's like there's this idea of like cool can we give people building blocks that let them kind of like build this parts of the software factory that they should own and like get really like skip all the of like dealing with you know what i mean of dealing with all this like sync infrastructure and stuff it's death by a thousand cuts you have all of this stuff that you're doing and especially what you were talking about if you're at a certain scale yeah then somebody has to do it you have if you're in an enterprise you for sure are going to be getting architecture reviews yep and somebody who is saying we can't do that because we need all of these things yep and you go ah yeah well our agent actually will will get to that I guess yeah but it's nobody actually wants to do that well it's like the people building these things like they have day jobs it's like you you do the skunkworks project I'm like who I built the software factory and like oh we have 30 users now we have 50 users and people are running a bunch of work through the through this background agent system I built and then people have feature requests and it's like do you want your job to become like maintaining this like side project you vibe coded or do you want to do your day job or like which which parts of it are worth investing in right same with github actions like in github actions you define how is the like language runtime setup how are the tests run but you don't build the thing that captures the logs you don't build the scheduler you don't build the thing that like catches a webhook from github and launches the work somebody else's labor Exactly.

45:21Dexter Horthy: There's like, it's like the overhead of building a CI. GitHub Action's got the interface really, really well. Like, hey, you build this YAML format, you commit it to your repo. I mean, Circle and Travis have this too. But like, basically that idea of like the hard, boring, annoying, not sexy part of this is the like wiring everything together and, you know, running things on demand and all of this. And that is the part that like, I think we want to be solving. And like, we've built the product to be like, cool, you Voltron it all together and you get this desktop IDE thing that you don't think about but then you can like take the pieces there's like vertical slices or that you could like oh I just want artifacts I just want when the agent writes a plan file I want that to be available to everyone on my team I want to get a slack message that it was written I want people to be able to comment I want to be easy to get those comments back to the agent.

46:08I smell a reefer.

46:14Dexter Horthy: Amsterdam. Just in case one more. Does that feel it? Just in case. Hopefully you're marking the ones you like. Bro, just by the way, we should rent one of those boats sometime and we should do a podcast on it. Well, today we're going to do just that with Dex. The only catch is... I think I heard that fucking... You probably heard it. That's hilarious. All right, let's try again. Fucking sneeze. The sneeze made me break. I need to ring some bells. don't let any of these pedestrians in the way. Central Station. So many damn bricks. Doing some acid. Are we hot? Come on. Let's do it. Come on. Let's get into it.

46:54Come on. Let's get into it. Let's do it. Come on. Let's go do it. Come on. I feel like we got it. You feel like we got it?

From the publisher

Two people, one tandem bike, and a lot of near misses along Amsterdam’s canals. Dexter Horthy, CEO and co-founder of HumanLayer and the person who coined “context engineering,” climbs on the back to talk about a problem every AI-heavy engineering team eventually hits: code that runs today that nobody understands three months later.


The ride starts with an old joke from the Claude 3.7 days: ship the slop, it’ll be GPT 7’s problem. That joke is aging badly now that GPT 6 is out and still can’t design clean abstractions. From there it becomes a tour of how agents are actually used right now. Computer use checks in for flights and books government appointments. Two hours of raw footage became a rough but promising product demo through Opus 4.8 and the Palmier MCP while Dex was out for a run. And the best MCP tools turn out to be the ones that mirror what Claude Code and Codex were trained to call.


Off the bike, over pour-over coffee, the questions get bigger. Are browser agents a bandaid until everything goes headless? What happens to ad-funded websites and upsells when an agent does the shopping? Why does chat still feel like such a low-bandwidth way to work with models? And why is every principal engineer quietly building their own software factory, then getting buried in the sync, streaming and approval plumbing that HumanLayer wants to turn into building blocks, the way GitHub Actions did for CI?


Demetrios Brinkmann: https://www.linkedin.com/in/dpbrinkm

Dexter Horthy: https://www.linkedin.com/in/dexterihorthy


Timestamps

[00:00] Intro

[01:38] The GPT 7 will fix it joke

[03:03] GPT 6 still writes spaghetti

[06:21] If it runs, does it matter

[07:28] Computer use that fills out your forms

[11:57] Editing a demo with Palmier MCP

[13:16] Why most MCP servers are bad

[14:32] Mirror the tools models were trained on

[16:57] Hill climbing a demo while out running

[23:49] Are browser agents a band-aid

[32:23] Why chat needs richer interfaces

[40:30] Every engineer is building a software factory

[45:13] The GitHub Actions moment for agents

More from Agentic Conversations (formally mlops.community)

All episodes