Sandboxing, Agent Harnesses, and Agent Teamwork

Shahram Anver: [00:00:00] We had this, like, wild discussion in the office where I was talking to one engineer saying, "What are you using?" And he, and he's like, "Oh, I use Caffeinate." And then I think Willem walked in and he's like, "Oh, I'm on Amphetamine." Like, anybody else listening to this would be like, "What the fuck are these guys talking about?"

Shahram Anver: Yeah. But these are, like, legit apps that are there- Yeah ... to keep you awake.

Demetrios: It's

Shahram Anver: not a

Demetrios: bunch of junkies that are- ... talking about- Or

Shahram Anver: maybe

Demetrios: Where should we start? Which one should we do first?

Shahram Anver: Why don't we do something with you? Like, uh, I, I think you're pretty into the harness stuff, right? I am. Let's, let's go-- I, I think we can go from, like, the, like, tech, like the harness, like how agents are done, to the human. Okay. So we'll end on the human.

Demetrios: Hmm. Harness to human.

Shahram Anver: Harness to human. Let's try

Demetrios: it. How much harness is too much harness? And-

Shahram Anver: Hmm ...

Demetrios: what even do you feel like is needed [00:01:00] on top of the harnesses that you get out of the box?

Shahram Anver: I guess, like, how would you define harness? Like, how do you see it? 'Cause it's so open.

Demetrios: Yeah. I think about harness as all the stuff that you expose to the agent-

Shahram Anver: Mm-hmm

Demetrios: to make it perform better. Sure. And so this is, uh, I kinda recycled the blog post from that dude Viv-

Shahram Anver: Mm-hmm ...

Demetrios: from LangChain, and it's like all of the prompts, all of the skills, all the- Mm-hmm ... if you have, like, logs, if you have the files, file systems. The model itself is just one thing, or the agent itself is one piece of it, and then you have everything that you add on to the LLM that helps it perform better.

Shahram Anver: Uh, you know, Claude Code, Codex CLI, these are harnesses in themselves, right? Mm-hmm. And then you'll have, you know, like, I think Viv and [00:02:00] Harrison, like they've got their own harness. I think OpenClaus, Py, if I'm not wrong. Yeah. There's a whole bunch of them, and they're, and they're, I guess, like different levels, levels of abstractions.

Shahram Anver: And I'd say like the Claude Code, Codex CLI is probably the, the, the most serious level of abstraction, right? Like it's the highest- Yeah ... where you, you-- it's just, you know, clearly you're kind of just giving it instructions. But then you can build even more on top of that-

Demetrios: Yeah ...

Shahram Anver: right, before you actually reach the user.

Shahram Anver: So that's why I'm always, like, curious, like, where that, like, definition is. So if we're saying the harness is basically everything-- Are we saying that? Are we saying it's between, like, the user and, like, the model, like everything in between?

Demetrios: Yeah.

Shahram Anver: Um, and then what's-- So in that case, um, I don't know how much it's too much.

Shahram Anver: Like, the way I, I think about it, it's like the job of the harness in this case is really to do th- two things, right? Like, agents really like fast feedback loops.

Demetrios: Huh.

Shahram Anver: And agents have, [00:03:00] I guess what you could say is, like this tendency to do really dumb things as well. So you almost wanna protect it from itself, and- Okay

Shahram Anver: you want it to have this fast feedback loop, so when it does do something dumb, it can react quickly and adjust, right? And know, like and know what it needs to do to adjust.

Demetrios: And when you say protect it from itself, you mean protect it from an agent that is doing the wrong thing? Yes. Or-- 'Cause I've also-- I was literally just having a conversation at lunch with my friend Mark about Protecting agents, like sub-agents from stepping all over each other.

Demetrios: Sure. And so there's many ways you can protect it from itself- Right, right ... whether it's isolation or- Right ... it's just like-

Shahram Anver: That's true ...

Demetrios: giving it the right tool at the right time.

Shahram Anver: That's true. That's true. So like, um- So it's like a really, really simple one. Um, I remember in the early days of Cleric, we were extremely, uh, careful about how we wrote tools, and [00:04:00] so we didn't give it acce-- So you imagine Kubernetes, like we didn't actually give it access to the kubectl

Demetrios: CLI.

Demetrios: Mm-hmm.

Shahram Anver: We wrote custom tools that were extremely rigid, so the agent could just do the thing that we wanted it to do and nothing else. But then, you know, models get better, and then very soon, like, you know, six months later, that turns out to be a really dumb idea. But at the, at the time, it makes a lot of sense- Yeah

Shahram Anver: right? Because, like, you're actually preventing it doing more dumb things.

Demetrios: Mm-hmm.

Shahram Anver: But now, like once y- you know, today, these things are just so good at composing multiple tools to get to the answer it wants, and then you get to a point where now the harness is like the thing that's restricting you from getting to, like, you know, better performance.

Demetrios: That's why the big question that I have is how much harness is too much harness? And, mm, like also how much is if you're trying to really map out what you want-

Shahram Anver: Mm-hmm ...

Demetrios: step by step and forcing it into one path-

Shahram Anver: Mm-hmm ...

Demetrios: there's going to [00:05:00] be times that you wanna do that, but then there's other times where you don't need to, and so by doing it, you're kinda wasting your time.

Shahram Anver: Right. Which is why I think it's an evolution, right? Like, you know, uh, eventually, like at the current state of the model and the current state of like whatever harness you're using, you-- whatever you're doing, if you're building an agent, I, I hope that you're monitoring these things and you're having these traces to see like where it's screwing up and where it's not, right?

Demetrios: Yeah.

Shahram Anver: So, you know, like we have-- we document failure modes. Like every time it does an investigation, we do a review, and we see what did it do well. But mostly, like we look at where it

Demetrios: went wrong.

Shahram Anver: How it did screw up. So we, we, we, we, we take the trace, and we put it through another LLM, and we have documented like labels.

Shahram Anver: Mm-hmm. Um, right? And then we try to figure out like, you know, did you spend too much time? You know, like for instance, like did we run this log query that just flooded the context window because we, we went too wide, right? Or we tried a [00:06:00] query that, you know, took six attempts when it should have just taken one because we were just experimenting.

Shahram Anver: Uh, right? Now, the thing that you don't want is to be too restrictive, but then when you notice that it's constantly doing these similar mistakes, then that becomes more of a design question. Like, is that something that you build into the harness, so you almost protect it from itself- Yeah ... like I said? Or is this like it's okay?

Demetrios: Yeah.

Shahram Anver: Right? Like, it's, it's not, it's not so fundamental that it's gonna screw up because if you try to overfit and you try to fit-- fix that, you're gonna screw up something, you know, elsewhere.

Demetrios: And like you said, if it gets that fast feedback loop and it's making the mistake, but it understands it- Exactly

Demetrios: can it fix it and correct- Exactly ... course correct in flight?

Shahram Anver: Exactly. So to me, it's almost like a game that you're playing where the model is just getting better and better at reasoning, and then it's getting a lot more capable at handling more subjective kind of tasks like this. Mm-hmm. Because obviously, like what you don't want is to be very deterministic, where it's like- Yeah

Shahram Anver: one size fits all. [00:07:00] Anything beyond like the most basic use cases, like SI is probably one of the worst ones because it's so dynamic and so complex. Yeah. Um, you actually want there to be some play, right- Mm ... for the agent. Like s- mistakes are okay. Like 100% perfect, um, like investigation is very suspicious.

Shahram Anver: Like it means that- Yeah ... either you like, you really overfit to that like one thing, which is like weird, or it was just so simple that you probably shouldn't, you know, pay attention. But now if it's like, you know, like 20%, it's just not getting anywhere, it's just constantly getting stuck, now you have a problem.

Shahram Anver: Mm-hmm. And then you could-- then I think that makes sense for it to be sort of a harness level, uh, you know, evaluation, let's say. Maybe the way I would frame it is that not everything has to be an agent.

Demetrios: Yeah.

Shahram Anver: Right? Um, like maybe an example that I can think of is, you know, in a, in a CICD pipeline, we go-- we, we have-- actually, like we try to do a, like a security review and a, uh, like a code [00:08:00] conventions review, things like that.

Shahram Anver: That does not need like a hundred different tool calls. It doesn't need to be like a very, very complex thing. It's pretty simple. Effectively, you're trying to see, okay, like what's the diff that's trying to be shipped?

Demetrios: Uh-huh.

Shahram Anver: And then what is some objective metric of what good looks like that you're passing to this LLM flow?

Shahram Anver: And then you just wanna know, is it pass or fail? And if it's fail, why? That sounds pretty deterministic to me. It's got like some LLM involved, and so maybe that's like, you know-- So that's where maybe a good example of where you don't have to overthink things and have- Yeah ... this really fancy harness and really fancy agent.

Shahram Anver: Could probably like get something off the shelf and, like, it, it'll work pretty well. Whereas like, you know, if you're trying to do, you know, like a coding agent or like an SRE agent, yeah, you're not gonna get away with something as, you know, simple as that.

Demetrios: Talk to me more about the evolutions that you had to make as the models got better within Flare.

Demetrios: Oh, that's

Shahram Anver: interesting.

Demetrios: Like you were saying, you know, we, we used to do this, [00:09:00] but now, a, a few months later even, we realized we don't need to do that anymore.

Shahram Anver: Yeah, these are always fun when you talk about like all the ways you screwed up.

Demetrios: Yeah.

Shahram Anver: Right? Instead of talking about like, "Oh, look at, look at, look at how awesome it's done."

Shahram Anver: Um, I, I think what like, you know, when we started, I remember like we had this like big debate. This was like, you know, late twenty twenty-three, and we went really fancy. Um, I remember talking about an architecture of like an ant colony where, you know, you have a queen that's talking to its workers, and its workers have like very distinct jobs.

Shahram Anver: So, you know, in this case, like the queen would be the incident commander almost, and like it would just fire off all these ants to go get, uh, data. And I mean, I think we just spent a little bit too much time on that. Like it seems-

Demetrios: Trying to create that architecture?

Shahram Anver: Yeah, because like there was clearly a time when, you [00:10:00] know, just, just for context, this was GPT 3.5, 4 just came out.

Shahram Anver: Mm. I don't think Anthropic was really-

Demetrios: In

Shahram Anver: the game ... seriously on the scene. Yeah. Uh, at, in the scene at that point. Maybe they were, but like definitely everybody was like really, uh, um, around the OpenAI models, and they just weren't very good. Um, I remember-- I, just for fun, I looked at our Git history, and I looked at the pro, like the prompts.

Shahram Anver: Like we were screaming, you know. It was like-

Demetrios: All caps.

Shahram Anver: Yeah, like, "Don't do anything but JSON." Like, "Just please return JSON." Like, "Do this, do that." It's like all caps, exclamation marks. I don't know if we-- like there were expletives, but I wouldn't be surprised. And then, you know, you're coupling that kind of, kind of, uh, you know, basic usage of the model, and you have such like, you know, sort of expansive architecture, something has to give, right?

Shahram Anver: So then we rebuilt everything. Uh, like it took us like, I don't know, like took, took us at least like three months to just realize, okay, like this is not gonna [00:11:00] get us anywhere. Um, and then we made it a-- we, we made it really simple, and I think that was good.

Demetrios: Uh-huh.

Shahram Anver: Um, you know, we call it a query agent, uh, where it was just like, okay, like, can we just have something very reliable where I give you a question, and you figure out where to go and just answer the question?

Shahram Anver: So it's not like ten workers, no supervisor, blah, blah, blah. It's like a single agent. And surprising, like I, I think that's, that's actually like CloudCode was amazing because it was also actually very simple. Um, and that was-- I'm not saying like we got simple first, but like we went really uber complicated, and then we went simple, and I thought that was like really effective.

Shahram Anver: The, the other piece was like, obviously like the tools thing that I talked about, right? Like we-- It was almost like a thing of if you know what you-- the problems that you want to solve, if you fix the tools, we felt [00:12:00] that that was a good way of getting this thing to be as deterministic as possible where we wanted it to be deterministic.

Shahram Anver: Mm. And so then let the, uh, the model actually figure out like how to, uh, you know, compose its way to the answer But I think, you know, eventually, like what, uh, you know, Claude Gods of the world did was say, well, actually the models have gotten good enough that you don't have to fix, you know, the, the tools anymore.

Shahram Anver: You can just write Bash, you can write Python and run that. And so what we had was basically relaxing that constraint and then moving, right?

Demetrios: And this again is like, uh, what I'm hearing is you're saying, "Oh, we just were giving-- like the harness was-

Shahram Anver: Exactly. -

Demetrios: less deterministic. We weren't so on top of it."

Shahram Anver: Exactly. But then the thing is like you always pay a cost, right? Because when you have tools that are that [00:13:00] deterministic, security is not a problem.

Demetrios: Mm-hmm.

Shahram Anver: Because you know off the bat exactly what the agent can and cannot do.

Demetrios: Yeah.

Shahram Anver: The moment you give it Bash and Python, now things get scary, right? And then so we couldn't really make that switch until we had to just sandbox the shit out of this agent.

Shahram Anver: Mm-hmm. Right? Um, so the way, at least like my mental model has shifted over the years is just- You want to have as much nondeterminism inside the box, but you wanna make a really strong box.

Demetrios: Uh-huh.

Shahram Anver: So that it can't-- Like as much as you can, it can't-- You, you wanna avoid breaking

Demetrios: out. I think you guys were some of the first folks that I heard talking about sandboxing.

Shahram Anver: Mm-hmm.

Demetrios: And you were creating environments too- Mm-hmm ... which is another one that became very popular-

Shahram Anver: Mm-hmm ...

Demetrios: because you had to simulate cloud environments- Mm-hmm ... to see if the SRE agent was running, right?

Shahram Anver: Yeah. Yeah. Like it's, uh... I, I think probably [00:14:00] the most complex thing that, at least in our domain, is that every environment we would go to is so different.

Demetrios: Mm-hmm.

Shahram Anver: Right? Like, I mean, I use myself as an analogy. Um, you know, I used to work at Gojek running the ML platform, and I'm working on problems like, you know, how do you make model inferencing for pricing, driver allocation, uh, you know, as, as efficient as possible. It's all on Kubernetes. But then everything I just said is gonna sound very familiar to someone at Lyft, at Uber, any, any, any kind of ride-hailing, food company, but it's not like I can join them today, and I'm gonna be productive.

Demetrios: Yeah. '

Shahram Anver: Cause they've got their own way of doing something. Like, yeah, the high level problem statement is a true, is cur-- is, is the same, but the way they do it is so different. And so, you know, if you assume that I am the agent, and I've gotten really good at these tools, like Kubernetes and all this stuff, I'm still gonna suck-

Shahram Anver: when I move into Uber, right?

Demetrios: Yeah.

Shahram Anver: And so we-- the way we [00:15:00] tried to solve for that, I know, like I, I think you had a really good chat with Willem and Shreya about this, right? A few, uh, what, what was it, like a year ago? Yeah.

Demetrios: It was probably a little more than a year ago now.

Shahram Anver: Right.

Demetrios: Yeah.

Shahram Anver: I said this was the primary problem.

Shahram Anver: Like, how do you test very quickly, um, across diverse environments and know that your agent's actually up to the task?

Demetrios: Yeah.

Shahram Anver: And so then, like im-- Like it's, it's been like a really interesting phase. You know, you're talking about like the evolution of Cleric. The investigation agent has just become the least important part of the stack.

Demetrios: Oh.

Shahram Anver: It's almost like the easiest thing to build now because that is like not what's actually durable. It's, it's like these environments. Like, how can you test? And how do you know it's working? Like, that's durable.

Demetrios: Mm.

Shahram Anver: Like the learning stuff, you know, we can talk about, um- It's kn-knowing-- almost like monitoring it and verifying whether it's correct.

Shahram Anver: This becomes, like, far more harder and more durable [00:16:00] challenges. Whereas, I think, you know, you can get your, uh, cloud code connected up to a bunch of MCPs and say, "Hey, go figure out what's wrong," it'll probably do fine.

Demetrios: Yeah. Because, uh, the search problem is much less complex than that. Is this working? Like the verification problem, I guess.

Shahram Anver: Right. Right. Right. Like once you go from beyond a simple kind of task where it's like, you know, something's wrong, it's ironic. Like the more com-- the, the more serious the problem, it's almost like there are some cases of that where it's actually very simple because it's serious, because something is down.

Demetrios: Uh-huh.

Shahram Anver: So it's quite binary, right? So y- it's, it's pretty obvious this is the problem.

Demetrios: Yeah.

Shahram Anver: Right? But like the more subtle it gets, now things get hard, right? Because now it's like very idiosyncratic. So that's where it's like really hard to know. That's where the verification comes in, that's where the learning comes in.

Shahram Anver: Because like you need to actually, like, look at the problem through the [00:17:00] lens of that company, like their conventions- Mm. -the way they do things. Because if you just come in with just world knowledge, it's gonna be really complicated because like the kind-- the, you know, out of maybe a hundred different reasons this could have failed, you have no way of knowing which ones you should be paying more attention to because they're more likely to be the cause, which an engineer- Yeah.

Shahram Anver: -who spent ten years in the company would immediately do, right?

Demetrios: Uh-huh. Which at the end of the day is kinda still the search space.

Shahram Anver: Sure.

Demetrios: But-

Shahram Anver: Exactly ...

Demetrios: you have to verify the- Exactly ... search in a much different way than... And I like what you're saying there, like the-- if it's a huge problem, you know, oh, the database blew up.

Demetrios: Or you know, it's n- Right ... normally pretty clear-

Shahram Anver: Right ...

Demetrios: if it's a very var-- if it's SEV one or whatever incident. Right,

Shahram Anver: right.

Demetrios: You're like, "Ah, we understand." But when it's just something that you're like, why is latency spiking by twenty percent here? What is going on? [00:18:00] What-

Shahram Anver: Those are the worst.

Demetrios: Yeah. And so then you really have to dig in, and you have to verify through that whole-- the search space again and have the agent know if it's on the right path.

Demetrios: And so that's been where you're focusing the attention now. You've-- and figuring out the sandboxes, figuring out the environments, figuring out how to verify.

Shahram Anver: Yeah. Like how do you just relentlessly and ruthlessly keep getting the accuracy up, and then as you're doing that, you keep falling into more and more of this subtle territory.

Demetrios: Hmm.

Shahram Anver: Then you need to build more stuff-

Demetrios: Yeah ...

Shahram Anver: to make sure that you're on the right track, right? Because y- once you've moved away from simple, that's where like all this other stuff that we talked about matters. I think- It's almost like th- that's why I, I'm, I've started thinking a lot now about just sort of the human agent interactions now.

Demetrios: Mm.

Shahram Anver: Um, right? Because, you know, in a world where the agents, you know, something [00:19:00] like 30 to 40% correct Uh, I actually-- I mean, this is a little bit of a detour, but, like, I was, I was talking to somebody, like an engineer who said, "I almost like it when the agent's always wrong because now I'm suspicious, and I'll check."

Demetrios: Oh.

Shahram Anver: Right? Whereas, like, if it's, like- That's right ... mostly right, I start trusting it. And then you- And then that one time, it like... So, so your spidey sense is now-- And I'm, I'm, I've, I've just seen the spidey sense just continuing to get challenged because, you know, like in the-- i-in, you know, w-well, all the stuff I was talking about with Cleric in the early days, you know, when we're doing these tools and things like that, you know, what's not said, it's like worth just being explicit about, is that, you know, we weren't very effective outside pretty simple kind of problems, right?

Shahram Anver: I mean- And you didn't

Demetrios: trust it at all- We--

Shahram Anver: Yeah ... I

Demetrios: imagine.

Shahram Anver: Exactly. Like, unless it was, like, very obviously the problem and we found it, um, you, you, you, you know, you didn't see that trust barrier being... So that [00:20:00] was, like, our focus for like- Mm-hmm ... you know, it's been like two and a half years of just, like, trying to get to that point where, like, it, like, you know, you're just implicitly trusting this thing.

Shahram Anver: And that's been happening, but now it's creating new problems, right? Because when that happens, people are now trying to throw you into more and more complex areas. Like, they wanna, like, give you more work.

Demetrios: Mm-hmm.

Shahram Anver: Um, and now it's getting a lot more subjective on, like, what's, what's correct and what's not.

Shahram Anver: And then people are worried about, well, you know, like, uh, how am I gonna keep learning, um, if this thing's getting so much better? Um, right? And so that-that's, that's what I find really interesting. That's where you're saying the

Demetrios: human aspect comes in.

Shahram Anver: Exactly. You

Demetrios: have folks that are starting to be, like, suspicious of it, but also they're recognizing that, well, if I'm always throwing it at all of my problems, then am I outsourcing the thinking?

Shahram Anver: That-- [00:21:00] But also, I think it's, like, more of a-- this shift that people have gone through, which is, you know, say, say twelve months ago, people weren't-- I don't know if people were actually that, like, concerned. Like, outside people who are, like, really deep into AI, I don't know if engineers thought, "Wow, these things are really good."

Shahram Anver: I, I'm not sure I, I, I, I heard that more and more.

Demetrios: No.

Shahram Anver: But we're clearly at a point, like, I think people just went, went into Christmas break, played around with, like, all this stuff, and they were like, "Holy shit," like, "things have really changed."

Demetrios: Yeah.

Shahram Anver: And now you can just see this trajectory that, you know, whatever it's not good at now, it's gonna get there.

Shahram Anver: Mm. Like, I know it is. And so now all of us are kind of kind of thinking like, "Okay, so what's next?" Like, what does-- what, what happens now where I'm not needed anymore, uh, in this one task?

Demetrios: Uh-huh.

Shahram Anver: Right? So imagine, like, you know, six months ago, you're talking to Cleric. Maybe Cleric gets sixty percent [00:22:00] there.

Shahram Anver: And now you have to, like, have two more follow-ups and get to their answer, and then you're done. But now it's going sixty-five. Now it's seventy, seventy-five, eighty. It's just keeping on going. And then you're like, "Okay, so what does this mean?" Right?

Demetrios: Yeah. Where do I add value?

Shahram Anver: Yes. And that's what I'm really interested in.

Shahram Anver: And so, so part of it is what you said, which is, like, the learning loop. But then also I think it's more, you know, like a more philosophical question, right? Mm-hmm. Like, where sort of am I? So in, in my mental model, it's like, I mean, I don't know. I, I don't-- I'm still working on, like, what the right analogy would be, but it's almost like all of us are gonna become managers.

Demetrios: Mm-hmm.

Shahram Anver: I don't think we're gonna... Like, you know, um If you look at sort of the w-- the kind of the online chatter, and like this is even happening in our, in, in our team, people are token maxing. People are like worrying about, how do I make sure my agents have what they need to get work [00:23:00] done?

Demetrios: Yeah. Like- How, how does it continue through the night while I sleep?

Demetrios: Right. It still is working.

Shahram Anver: Right. Like I had this like... We had this like wild discussion in the office where I was talking to one engineer saying, "What are you using?" And he, and he's like, "Oh, I use Caffeinate." And then I think Willem walked in, and he's like, "Oh, I'm on amphetamine." Like anybody else listening to this would be like, "What the fuck are these guys talking about?"

Shahram Anver: Yeah. But these are like legit apps that are there- Yeah ... to keep your, uh-

Demetrios: It's not a bunch of junkies that are- ... talking about- Or

Shahram Anver: maybe.

Demetrios: Yeah.

Shahram Anver: But, but yeah, like so, so then, uh... But you know, if you just like double-click a little bit beyond like the sweatshop vibes- ... um, it's basically it's like, "Hey, I have these workers, and I kind of care about making sure that they're okay.

Shahram Anver: I wanna make sure that they're not gonna be interrupted- Yeah ... they've got what they need- Sure ... to get stuff done." I mean, it's pretty kind of a managerial problem.

Demetrios: Yeah.

Shahram Anver: Right? It's- It's

Demetrios: just the managers don't tend to give humans [00:24:00] amphetamines-

Shahram Anver: No ... unless- Well, well not-

Demetrios: Yeah.

Shahram Anver: At least not all of them. Like the ones we-- you, you should watch out for.

Shahram Anver: Those are the ones that-

Demetrios: Exactly ... um- Yeah. So but the, uh, apart from that, I agree with you wholeheartedly because when we had the coding agent conference- Mm-hmm ... a month back, I remember one of my friend's bosses, who is their like VP Eng at a fairly big company. Mm-hmm. Like, uh, he was so excited because he went to the workshop-

Shahram Anver: Mm-hmm

Demetrios: and he was like hands-on shipping code, like merging PRs, super excited. And then when I wrote about that on LinkedIn, people were like, "Yeah, well, why wouldn't they be?" Managers have been doing this. Like what do you think a manager is? It's almost like a vibe coder. You're- Right ... managing agents or you're managing people.

Demetrios: It's just people are actually kinda harder to manage.

Shahram Anver: Hundred percent. Exactly.

Demetrios: Humans are [00:25:00] very messy. Agents, less so. You can really grind and give them the amphetamines, and it's all right.

Shahram Anver: Yeah. Or you can just be, um... Yeah, there's-- as long as you don't need to be too nice to your agents.

Demetrios: Yeah. Polite. They don't ask for raises every couple months- No

Demetrios: uh, or more shares, all of that stuff. Or they don't like complain about sitting next to somebody that smells bad. All of that stuff that humans have to deal with.

Shahram Anver: Right.

Demetrios: You don't deal with that and-

Shahram Anver: But it, it's funny to me because it's, it's like I, I think people have gotten really caught up with how addictive it is, right?

Shahram Anver: And I think that's coming from this place of You know, like, I think if you just went through and got to like, you know, eventually like the end state, it's a very stressful job being a manager, right? Like, you're constantly like, "Hey, I've got, you know, 10, 20, 30 people depending [00:26:00] on me, and what I say we're gonna do- I've got thirty times, like, the bandwidth of that week assigned to this problem that I've said we have to do.

Shahram Anver: And now, you know, like-- And this problem goes from CEO to engineering manager.

Demetrios: Mm-hmm.

Shahram Anver: It's pretty freaking worrying, right? Like you-- And so, like, what I'm seeing with the amphetamine and, like, all these conversations is, like, the, the kernel is that same worry. Like I, I've got all this, like, bandwidth of tokens.

Demetrios: Mm-hmm.

Shahram Anver: Am I spending it correctly? Like, am I doing the right thing? I-- Some of us aren't. We're just, like, shipping all kinds of shit. But, like, generally, like it's, it's just gonna end up there, right? And, and then you're back to being this, like, person who's like, "Okay, like, here's-- I've got X resources."

Demetrios: Mm-hmm.

Shahram Anver: "I've got a problem to sort out. How do I, you know, corral my army of people or, like, agents to actually go, you know, solve this problem?"

Demetrios: Yeah, make sure we're all rowing in the right direction.

Shahram Anver: Exactly.

Demetrios: And taking [00:27:00] advantage of every moment because any moment that the agents aren't-

Shahram Anver: Right ...

Demetrios: looping.

Shahram Anver: Well, hopefully we don't get too stressed out about that because that, that is a, like a dark hole-

Demetrios: Yeah

Shahram Anver: I, I think. But yeah, exactly.

Demetrios: You don't have eight agents constantly running, and you feel like you're-- What are you even doing, right?

Shahram Anver: I know. I mean, I feel that. Like I, I, I've-- I, I don't know the last time where I haven't... So, so we-- I'm sure, like, people are using their own terms, but like internally, we use the term like the hopper.

Shahram Anver: It's like you're cooking, like- Mm-hmm ... well, I'm just gonna put another one on the hopper. And so I, I, you know, I, I need-- I feel like I have this need to put something on the hopper before I go to bed.

Demetrios: Mm.

Shahram Anver: And I have just infinite satisfaction waking up to see it's, like, almost done. Yeah. I think they're pretty cool with, with the harness discussion and all that stuff as well, because, like, you-- what you're implicitly doing is every time it screws up, you're kind of thinking, "Oh, what should I have given it?

Shahram Anver: Like, what could I have given this?" Like, you know. And it's, like, really what [00:28:00] managers do, right? Yeah. Like, you know, I, I was, I was reading a post from-- I can't remember who, but she was like, the first thing she asked on her one-on-one was, uh, like, "Well, how-- what's your work setup like?" Mm-hmm. "And can I fix something?

Shahram Anver: Like, are you comfortable?" I mean, are we literally doing that for agents now?

Demetrios: That's the harness.

Shahram Anver: That is the harness, exactly.

Demetrios: Uh, what I've been seeing, and I imagine this is going to be something that maybe it's fleeting, maybe it becomes more prevalent, is that you have to do a lot of the upfront work-

Shahram Anver: Mm-hmm

Demetrios: in order to be able to put something on the hopper and let it run.

Shahram Anver: Yes. Yes. But by upfront work, I assume you mean, like, you know, speccing it out properly. The

Demetrios: scaffolding around it. You can't just kinda like willy-nilly throw in one or two sentences and be like, "Go and have fun."

Shahram Anver: Well, you can, but then you're just gonna have like a pretty shitty, uh-

Demetrios: Shitty outcome

Shahram Anver: uh, exactly. I mean, that's where I guess like cost and things like that come in. But yeah, so, so what, [00:29:00] what, what I've-- what I, I'm thinking is like, okay, so right now we're kind of like, you know, some of us more than others are using this more and more, like effectively.

Demetrios: Mm-hmm.

Shahram Anver: And then, you know, what are these people doing that's more effective, and how are we learning from that?

Shahram Anver: And like, where does this all go? Mm-hmm. Like your point, um, you know, this whole spectrum of development and the Ralph Wickham loop and all this stuff, like it's coming out because like, you know, these are-- like Dex was amazing at the coding agent- Yeah ... conference, as always. Uh, like it's all like these lessons learned from people who are just like doing it all the time, and we're all trying to learn from it.

Shahram Anver: Um, but I guess like my, my, my thought here is that I think it's kind of like you wanna be a mix of a manager and a craftsman, right? Because- You know, in a world where you can build anything you want, effectively what is left is your taste and your judgment, right? And I-- That's, that's why I, I, I think there's a very dark way [00:30:00] of taking this, which is that, you know, you lose all agency, you stop learning.

Shahram Anver: Like you-- Like this thing- Mm-hmm ...is just gonna... And I, I don't know. I guess, like, if you take anywa- anything from this conversation, I hope that we can have a more positive view of it, which is that, like, actually, the really fun stuff about building product is putting your personality and, like, your thoughts and, like, your view of, like, what you wanna bring to the world out there.

Shahram Anver: Now you can get there much faster.

Demetrios: Mm-hmm.

Shahram Anver: And so, like, how do we build for a world like that, where that becomes, like, all the-- wh-where you spend all your time on? Whereas, like, the-- So the toil of getting there is something what the agent is doing.

Demetrios: Mm-hmm. It goes back to the, those feedback loops you were saying before.

Shahram Anver: Right. That's what I was thinking. Like i-if, you know, if you-- 'Cause the thing with-- I think the managerial word, which I don't like as much, is that it almost suggests that you don't know much about the craft. You're just focusing on- Yeah ...the sort of puppet mastering aspect of it- Yeah ...which I, you know, I-- that's why I don't like the [00:31:00] word.

Shahram Anver: I like the whole craftsmanship thing because, like, it, it, it, it's about mastery, right? So you're almost, like, in charge of a workshop, and you're trying to get something done. And what you don't want is to judge based on what's produced. You actually want to know what's going on.

Demetrios: Mm-hmm.

Shahram Anver: Right? You, you don't want to be abstracted away too much.

Shahram Anver: So I think, like, even with agents, I think the mistake would be to say, "This thing's so good, so we're just gonna wait for it to finish all of its code, and then I'm gonna review all of its code." I'm not sure that that's effective.

Demetrios: Well, you did wanna say something ab-- uh, before we hit record about how you wanna be able to see as it's making its decisions, what it's making- Sure ...and then steer it.

Demetrios: And that steerability, I think, is a little bit lacking-

Shahram Anver: Mm-hmm ...

Demetrios: sometimes.

Shahram Anver: Mm-hmm.

Demetrios: But also, if you're gonna try and scale to running eight parallel agents- Mm ...you just can't expect to steer- No ...eight agents in [00:32:00] parallel.

Shahram Anver: That's true. But that's why I think the point of abstraction is, like, useful, right? Because, you know, if you've got eight agents and you're writing some, you know, big project, you know, it's gonna produce just reams and reams of code.

Shahram Anver: You can't read all that code.

Demetrios: Mm-mm.

Shahram Anver: But an abstraction above that is what are the decisions that it took to get to that code, right? Mm. Like a really dumb example is You know, an agent is probably gonna be really good at writing a macOS desktop app. Um, it's gonna get, like, really good. It's gonna look beautiful.

Shahram Anver: But maybe the judgment is that we shouldn't write a desktop app. It should be a web app.

Demetrios: Yeah.

Shahram Anver: Um, right, because I want more customers, and, like, uh, I, I, I think there's a bunch of people on Linux who's gonna like this stuff as well. I mean, that's a judgment call. And so then there's a decision trace somewhere where your agent or your product manager agent is like, "We should build a Mac app."

Demetrios: Mm-hmm.

Shahram Anver: And so then, you know, at that [00:33:00] point, you could probably go, "Well, no, we shouldn't. Um, I think we should build X instead." That is, I think, infinitely easier to review than- Than the code that was written ... all the code that was written, right?

Demetrios: But isn't that just making you the bottleneck?

Shahram Anver: Well, that's the thing.

Shahram Anver: Like, I think it-- if, if you become the bottleneck, then it should be for a good reason. You, you are actually, like, putting-- inserting yourself in there. I think there's the kind of two constraints. Like, one is your test. So hopefully, like, you know, you are operating at that level where you're injecting a test and that ideally is like something you're exercising and you're getting better at.

Shahram Anver: And the second one is like, I think the real world also has constraints, right? Like in a world where everything gets amazing and, like, you can produce infinite code, your customers can't take that much change anyway.

Demetrios: Mm-hmm.

Shahram Anver: Right? Like, you can't have-- You know, like a lot of these companies and the services that we are used to took decades to evolve into what they are.

Shahram Anver: Like, you know, if I-- if you just [00:34:00] speed ran someone who uses Uber today to everything that's-- they've offered, you know, over the last ten, fifteen years, I mean, that's not possible, right? Yeah. Because people can't take it. So I think, yeah, it is a bottleneck, but, like, it's-- it only makes sense if it's adding value.

Shahram Anver: Mm-hmm. And I think those, those are two ways of adding value.

Demetrios: Well, that's a fascinating topic to think about is where you need the human-

Shahram Anver: Mm. -

Demetrios: to plug in because if it is-- and going back to your example of the desktop app versus a, a web app, that's one question that potentially you plug in a human for or you don't.

Demetrios: And it's-- Because maybe it just doesn't matter and you can build it, but, like, what decisions need to actually be outsourced to the human?

Shahram Anver: Mm.

Demetrios: You know, where do you plug the human in, and how does the model know that this is a really [00:35:00] big decision? Mm. Before I make this decision and go down this potential path, I should probably ask someone.

Shahram Anver: Yeah. I, I mean, I think it's, it's useful thinking about like what the end state-- I mean, there's no true end state, but like at least like I think in the next twelve months, like what, where I think we should be. It feels to me like you're kind of approving exceptions, right? Like when, you know, like say you have an employee and they're supposed to like, you know, do something.

Shahram Anver: Like for instance, like book this podcast studio, right? Mm-hmm. It's, it's, it's likely that you're giving them some guidelines, right? Like, "Here's your budget. Here's kind of the area that I want you to look at. And, you know, you go figure it out." But if you're gonna violate any of these bounds, and there's a good reason, then I need to know.

Demetrios: Mm-hmm.

Shahram Anver: And I, I think it's kind of similar here.

Demetrios: But again, then you have to really upfront define those- Yes ... bounds clearly- Yes ... and put that scaffolding in place-

Shahram Anver: Yes ...

Demetrios: and then let it run.

Shahram Anver: But, but [00:36:00] I think that's the work which is valuable.

Demetrios: Yeah.

Shahram Anver: Right? Because I think those are the decisions which today, like, you know, live in people's heads.

Shahram Anver: There's a lot of implicit decision-making that's happening, right, which is not explicit. But it, it was okay, or at least it's okay in a world of humans, because when we, when we're not sure, you call a meeting, you send a Slack or things like that. But it just does not work with agents because agents work on agent time, right?

Shahram Anver: Like you-- it needs an answer right now. Like from an SRE standpoint, like where, you know, that's a domain, you know, I think about a lot. It's like y- how can you reliably get an agent to fix something at two AM if it doesn't have clear bounds on what it can do and what it can't? It just can't. It can't call a meeting at two AM.

Demetrios: Yeah.

Shahram Anver: Like we call meetings at two AM because there's an incident, and like you call everybody and like, "Guys, what do we do? Like should we restart this? Should we stop it? Should we do this? Should we do that?" Unless those decisions become, you know, preserved or like, you know, queryable, you're not gonna get [00:37:00] anywhere.

Shahram Anver: Mm-hmm. So I think the phase we're in now, it's all about actually, you know-- I think Jaya from Foundation Capital did a re-- amazing piece called Context Graphs- Mm-hmm ... which really just blew up. But I think it's like all around kind of a similar idea here where that's the valuable data now. Like what are the decisions that you know are okay?

Shahram Anver: Like what are the bounds? And then how do you actually track the decisions that the agents are making so that you're actually having this compounding loop of preserving- Yeah ... and it's, uh, these instructions, uh, in a way that humans can read them and, and like go, "Yeah, yeah, that looks fine. No, no, don't do this."

Shahram Anver: Or at least you're defining an exception. Like, "As long as you're over here, do whatever you want. But if, if you're coming out, talk to me 'cause I need to tell you what to do."

Demetrios: Are you guys doing that? Is that one of those more complex things that you started to add where you have decisions being made in Slack-

Shahram Anver: Mm-hmm

Demetrios: and you need to then incorporate that into the agent so that if it encounters this problem again in the future, it kinda has a [00:38:00] roadmap?

Shahram Anver: We're doing that. So that, that to me is like part of learning where you-- I don't think it's as advanced as I, as, as I'd like, but that's the path which I wanna be on. So, you know, I've, I've been railing on, um, you know, on social media about like this forward deployed engineer kind of model.

Shahram Anver: The reason I'm not bullish on that is that you're actually not building the thing I'm talking about, right? Like, what I want when you deploy Cleric is I'm giving it every opportunity to learn these rules And if you don't give it that opportunity, and you're just gonna throw engineers at the problem and try to, like, tune it, then you're, then you're not solving the actual valuable problem.

Demetrios: Mm-hmm.

Shahram Anver: Right? So to your point, on Slack, there's a wealth of information. Like when we get added to an alerts channel, I mean, there's a ton of stuff like, you know, when this thing goes, goes down, usually it's because of this other thing. And here's what, you know, over the last six months, our engineers have tried, and like [00:39:00] this is what generally works, this is what doesn't.

Shahram Anver: You can build a pretty effective map just from this data. It's, you know, it's like unstructured data. It's like it's, it's, it's very valuable, right? And then that can be replayed back to the human to say, "Here's what I learned. Does this look right?" Um- Yeah ... right? And then now to a human, that's actually like a very compressed form of knowledge.

Shahram Anver: Like these are sort of what I've learned, like what I should do, how your environment works. Does this look right? And to them, it's like very easy to just edit it.

Demetrios: Yeah. I do like the distinction too of how you're saying the real value is in that type of thing and being able to create these, uh, paper trails in a way or create that, like, decision-making process, the SOPs-

Shahram Anver: Mm.

Demetrios: And then doing the upfront work-

Shahram Anver: Mm ...

Demetrios: of creating those so that the agent can do the hard work of like actually- Yes ... implementing it.

Shahram Anver: Yes. [00:40:00] Exactly. And you, you wanna design your systems to make that hard work like valued, right? Like, like it, it has to happen. Like we've tried- Yeah ... a lot of times, like to do-- like in the SRE world, you'd, you'd insist on teams doing postmortems, documenting them- Mm-hmm

Shahram Anver: things like that. It's really hard because you're asking people to do something with no immediate tangible value. Like unless you've really got a good process down, and this happens in every industry, right? Like just document your shit. Yeah. Like just make sure you're writing stuff down.

Demetrios: Everybody knows that- Right

Demetrios: it's a good idea, but it's not necessarily-

Shahram Anver: Right ... exciting. But I think agents are super interesting because now you do get that feedback loop- Yeah ... very quickly, right? Like if I write something down or I tell the agent, "Hey, remember this," and then the next time something happens, it's like, "Oh yeah, thanks, D.

Shahram Anver: You told me about this, so I, you know, I did it this way instead." Now you feel, "Oh, okay. That's, that's cool." Right? Like-

Demetrios: So in a way [00:41:00] What you're constantly doing is updating documentation and updating

Shahram Anver: Pretty much, like just for the agent.

Demetrios: Yeah, just for the agent. Yeah. And, and I remember, I think Willem told me, yeah, normally you don't necessarily need to ingest the whole Slack thread- Mm.

Demetrios: -because by the end of a Slack thread, there's a decision that's made.

Shahram Anver: Mm.

Demetrios: And did you do it with the-- Is it the user experience that like at the end of the Slack thread, once you know what needs to happen, then you just like @Cleric, you know, like update this or here's the information? Or is Cleric just there reading everything?

Demetrios: Because I'm wondering how you don't have a lot of noise.

Shahram Anver: Well, that's a great question. Um, noise is inevitable. You will learn completely useless things.

Demetrios: Yeah.

Shahram Anver: Um, th-this will happen. Um, but I think to us, like it's more about curation, right? Like, how do you-- Like we call it rot. [00:42:00] I think it's a pretty standard term at this point, um, that you're just gonna have this sort of memories that you learn which aren't very helpful.

Demetrios: Mm.

Shahram Anver: Um, I mean, there's a couple of ways. Like one is we give people an explicit way to create them and forget them. So if it's like we totally screwed it up, we learned something we shouldn't have, or like we're somehow not learning something we should, then you can just take over the wheel and be like, "Yo, you need to know this."

Shahram Anver: Like, "Remember?"

Demetrios: So it's like, yeah, commit this to memory-

Shahram Anver: Yeah. -

Demetrios: um, or forget this-

Shahram Anver: Exactly ...

Demetrios: eternally.

Shahram Anver: Exactly. Yeah, like don't-- do not mention this to me again. Yeah. Um, but j-- like that's usually for extremes, right? Because people are lazy. Like you gotta design for that. Yeah. Like you, you don't wanna make people say, "Remember this all the time."

Shahram Anver: So we do implicitly, um, learn as much as we can, like from people's conversations. We also try to make it really cheap for people to, um, you know, [00:43:00] say good boy, Cleric, and bad boy, uh, like with ratings and things like that- Mm-hmm ... because then that also gives a signal that all the decisions that we've made so far, if it's a five out of five, it's pretty good chance that we've did the right thing, right?

Shahram Anver: So then we can look through it. Whereas if it's like a one out of five, even though we're like, "Oh, we did a really good job," then you know, "Hmm, okay, something went wrong." So either there's something we missed or like something we, you know, over-indexed on or something like that. Um, so that's-- that also helps a lot with the learning.

Demetrios: Back-- Continuing with the noise theme, I imagine you can Never be satisfied on all the data that you're collecting. Like you-- Whether it comes to, all right, you've got your Datadog, you've got your code, you've got the Slack conversations, you've got the postmortem transcripts, you've got everything and anything that you can get your hands on.

Shahram Anver: Mm-hmm.

Demetrios: Are there-- Is there a little bit of like a eighty-twenty principle where you say, "You know what? [00:44:00] If we just get the Slack conversation and the commit history- Hmm. -that gets us pretty far."

Shahram Anver: Yeah. Um, I'd say Slack's amazing, code's amazing, and logs are amazing. Um, these three get you very far. Um, Slack mostly for just historical stuff.

Shahram Anver: Um, and this is again, like with the assumption that you haven't been very disciplined in storing everything. Like, you know, there are some folks we work with who've done, done a really good job. So, you know, if you go to their Confluence page, it's like really well laid out. Um, like all their lessons learned are in a- Hmm

Shahram Anver: separate database. So then that's like obviously much better. Um, but generally like, yeah, like this kinda text-based knowledge, um, is helpful. The code is interesting because you're, you're looking for kind of like a source of truth kind of knowledge. Because no matter what you see on Slack or lo-- [00:45:00] or documents or things like that, you never really know how up-to-date it is, right?

Shahram Anver: Like, you know, I, I could say something today, and a commit tonight could completely invalidate that. Yeah. Right? Like completely change everything because we've done a big refactor. And so you still need to have that source of truth. So it's at-- Like to me-- to us, like code and Kubernetes, like InfraState is really helpful because the agent can just be like, "Okay, hang on, like, like what is true today?"

Shahram Anver: And then like, "Can I work backwards to, okay, what do I know historically?" And then like, how do I... You, you can't go the other way around. Yeah. Like you're trying to piece together a truth from history. It's hard.

Demetrios: Yeah. Now, the thing that Willem always tells me when I talk to him is like, it's pretty easy to get something up and running-

Shahram Anver: Mm-hmm

Demetrios: uh, in terms of an SRE agent that you can demo.

Shahram Anver: Yeah.

Demetrios: Like all- [00:46:00]

Shahram Anver: Hundred percent. Any ...

Demetrios: any agents basically. Yeah.

Shahram Anver: Could

Demetrios: put, get it

Shahram Anver: overnight- Yeah ... and then yeah, look, look pretty convincing.

Demetrios: And then he's like, "But dude, when you get into real production systems, it is so much harder to get an SRE agent going."

Demetrios: Can you explain like why is that? What's the difference there that is the production gap that you see?

Shahram Anver: I'd say it's two things. One is just the human element. Uh, the expectations are so much higher, right? Because, I mean, the way I like to think about it is if I'm-- if it's a coding agent and I got 70% of the code right, as a user, you know, you just saved me 70% of the keystrokes I would have to make.

Shahram Anver: I'm, I'm pretty happy, right? Like I, I can, you know, look through it. Um, even if it's not right, it's, it's kind of cheap for me to know that it's not correct. I can write some tests. [00:47:00] I can, I can get it over the line. I'm not, I'm not pissed off with you for being 70%. But we're on the other extreme, right? Like, you know, if I'm seventy percent correct, and I led you down this path, which is not exactly right, and you've spent twenty, thirty minutes in a pretty stressful situation sometimes, you're gonna hate me.

Shahram Anver: Right? Like, like forget about tech. It's just from a user point of view, like that's-

Demetrios: Yeah.

Shahram Anver: Right?

Demetrios: Like, why did you lie to me?

Shahram Anver: Right. And you know, and, uh-

Demetrios: That's actually, yeah ...

Shahram Anver: I like to think about it as like- It's

Demetrios: pathology ...

Shahram Anver: Yeah, yeah. You know, like the, like the personality of people who are in the, you know, the, the, the, uh, like the, you know, it's like the, the beast of the belly, right?

Shahram Anver: Like you're just the belly of the beast. Yeah. Of the belly. Um, you b- you, you, you get a certain kind of personality, right? Like you're, you're not gonna tolerate, like you're, you're very risk-averse. You- Yeah.

Demetrios: Adrenaline's pumping.

Shahram Anver: Right. 'Cause you [00:48:00] know, like ba- basically, the kind of companies we work with are larger companies because, like, if you're a startup, you don't really have a production problem.

Shahram Anver: Like you don't have that many customers. But once you have a lot of customers, you're gonna allocate a certain set of people to look after them, and these people are like, obviously, for very good reason, like very, very careful. But on the, uh-- then when you bring it to the technical side, that's also where the complexity is because, um, you know, for-- Let, uh, let-- I'll just take code again because I think most folks probably listening to this are using coding agents.

Shahram Anver: Most of the truth that you need is in your code, right? Like, like for, for a coding agent, you're actually going deep. You give it this one thing, your coding repository, maybe it's a couple, and you're saying, "Go build me this new feature." The complexity of the coding agent is the kind of inputs of the tasks you give the coding agent is potentially infinite.

Shahram Anver: Like, you can ask it to build anything you want.

Demetrios: Mm-hmm.

Shahram Anver: But the base it's doing it on [00:49:00] is finite, right? Like it's just got- Yeah. With SRE, it's the opposite, right? The, the answer could be anywhere.

Demetrios: Uh-huh.

Shahram Anver: So the-- there's no base. Like it's a, it's a, it's a horizontal kind of breadth-first search, but the kind of problems are kind of finite.

Shahram Anver: Like it's a latency spike. It's a, you know, something's down, a user's upset, there's a bug. Mm-hmm. You know, it's not like-- So the complexity really is, like from a technical standpoint, that like the-- it's a needle in a haystack. Like it could be anywhere. And so when you couple that with the user problem, hopefully that gives you a sense of why this is-

Demetrios: Yeah.

Demetrios: Flip that right on its head, dude. Yeah. Oh, man. One of the big questions that I am sure you think about a lot is like, are we just gonna end up becoming a skill? Like the, is this whole thing gonna get so good that now we're a markdown file that's part of a coding agent?

Shahram Anver: You know, [00:50:00] like my default mode, uh, is I'm just a very optimistic guy, so that's probably like a good thing to just spread out there.

Shahram Anver: Like, that, that's why I think Willem and I work really well because he's definitely a lot more like on the-- Like he'll see all the risks. Like you know him well, right? Yeah. So I see all the optimism, and then we're trying to like meet in the middle. Um, I think that's part of why we should be having this, this, this discussion, and like people like us should be talking about this more because like the whole point is to avoid that future, right?

Shahram Anver: Mm-hmm. Like that's why I think, you know, redefining how you see yourself as this craftsman, as this manager, and like where you're actually adding value, and just knowing where the value is and just going for that, that's what you gotta do, right? Because, you know-- I-- You know, just think about like there's two problems that-- obviously, there's many, but I'm trying to reduce it to two problems that the AI world is working on.

Shahram Anver: The first one is like, can you [00:51:00] produce output that's, you know, like problem-free, like no bugs, uh, it's, you know, what I'm saying is generally correct, or, or rather like it, it works. And the other one is like, is the work you're doing correct in that like is it directionally what, what you want? Like is it a should I build a macOS app, or should I build a web app?

Demetrios: Mm-hmm.

Shahram Anver: Like that takes subjectivity and taste. Whereas the first one is more like, you know, if I ask you to do a thing, did you do it, right? If your whole world is around the first problem, I mean, I've-- there's-- I don't have good news for you. Like, like- Yeah ... it, it is-- you have to assume that AI will just get so good at solving that first problem that you need to think about, like, okay, what happens with the second problem?

Shahram Anver: Which is like, you know, that's where the subjectivity and taste piece comes in, and the real-world constraints that you can't just build everything overnight. People can't take it. Then I think you wanna be like, "How do I be the purveyor of taste, and how do I be the one who's really good at, like, [00:52:00] making these decisions?"

Demetrios: Mm.

Shahram Anver: Because I have this thing that can actually produce working artifacts, but how do I make sure it's the correct working ones?

Demetrios: And continuing along those lines, do you ever feel like Cleric could just be abstracted into a skill?

Shahram Anver: Could be. I mean, I, I think, like, I think our, our job as founders is, like, to constantly think about what is a durable problem and what is a problem that you should not be spending an iota of time on.

Demetrios: Uh-huh.

Shahram Anver: Right? Like, like when we started, my, my big rule to everybody was like, "We are not fine-tuning models." Like, this is, like, ludicrous. Yeah. Uh, people will figure this out. They're gonna get way better at, like, building way better models. Just assume the models will get better. I know you hate screaming at the model with JS, like, "Please give me JSON," but this will be fixed.

Shahram Anver: Don't

Demetrios: worry.

Shahram Anver: [00:53:00] Like, spend time on the agent. And then our whole thing, like, last year was, like, the agent layer is commoditized. Like, it's a waste of time. Like, I think there's gonna be way better harnesses than we can ever build. There's gonna be, like, better agents and all that stuff. So then, you know, what we're thinking about is, like, very much on, you know, the verifiability and learning and things like

Demetrios: that.

Shahram Anver: Mm-hmm. I think the same time next year I'm gonna be saying something different. Yeah. Like, that's the point, right? Like, you don't build something that everybody else is gonna do better than you. Like, build a thing that's hard in your domain that's gonna last. And, like, I think the really tricky thing with AI is that that thing is changing every 12 months.

Demetrios: Yeah.

Shahram Anver: But then you gotta be primed to keep making that jump every time.

Demetrios: Well, talk to me about what 2027 looks like. It-- you kinda mapped that out in this blog post that you wrote.

Shahram Anver: Right. I was trying to think about, like, what this world looks like. Um, and- To me, it's almost like, you know, you're, you're, you're waking up and you've, you've had all these agents that have been working, you know, for you, and then they're sort of surfacing ex-exceptions.

Shahram Anver: [00:54:00] So maybe your coding agent shipped a change like to a config and something ended up breaking at two AM and your SRE agent notices this, reverses the change and watches what happened until, you know, the metrics seem to be fine. And then you-- when you're waking up, you're just getting a report that- Mm-hmm.

Shahram Anver: -you know, this is what happened. Here's the decisions I took. Like, you know, I noticed that the coding agent did X, so I decided that this was wrong, so I reversed it and so, you know, hope that's okay. Don't worry, like everything's fine. Or it could be like it's, it's about to make a mistake because, uh, it doesn't know, so it's like surfacing kind of, uh, like a question to you, like should I roll back because I don't know, this is uncharted territory.

Shahram Anver: But that's almost like to me the interface that we will be operating, um, with these agents. Mm-hmm. So to me there's kind of two interesting parts there. One is [00:55:00] I call it like the one-on-one problem, which is like as a human, like at what level am I doing my one-on-ones with my agents? And that's like the policies that we're setting, right?

Shahram Anver: Mm-hmm. So, you know, like it's okay to roll back if X happens. This is the way we build things, things like that. And then the other one, I call it like the stand-up problem, which is I think it's gonna get maybe a little bit more interesting where once you have teams of agents, I'm not sure we've done a great job yet as an industry to have these agents working together.

Demetrios: Mm-hmm.

Shahram Anver: Right? Like today, you know, Cleric's an SRE agent and I've just talked about how it's totally different to the coding agent, but it's so obvious that we have to work really well together.

Demetrios: Yeah, and how do you share context?

Shahram Anver: How do you share context? Things like that. How

Demetrios: do you work off of the same planning?

Shahram Anver: Right.

Demetrios: And you're able to cross one thing off or-- Yeah, that's-- that is a really interesting one with, w-with the proper isolation.

Shahram Anver: Right.

Demetrios: So it's a little bit tricky in [00:56:00] that regard because you want it to be isolated, but you want it to also share and share resources when needed and so it gets-

Shahram Anver: Yeah. It- -

Demetrios: messy.

Shahram Anver: I mean, I, I think I think what's been really interesting about just thinking about agents is that there's so much overlap with, like, how we've solved the same problem for humans. Mm. And like, you can borrow a lot of that stuff. Like, I don't think we need to reinvent the wheel. Um, in this case, you know, y-you forget about agents.

Shahram Anver: Like, I've got a product engineer. A product engineer's main motive is to increase the value of the product to their customer. They're not-- I mean, hopefully they are, but they generally don't think about security and reliability. Mm. They're just like, "Ship this as fast as I can, get my promotion." And then SRE is like, "Whoa, what are you doing there?

Shahram Anver: Slow down. Like, this, this could break something, and that's gonna be bad." Security will have the same thing, right? But then, like, how do they work together? So, you know, there'll be something like if there's a big change, maybe the product engineer will be like, "Hey, I'm shipping a [00:57:00] PR. Can one of you guys have a look and let me know what you think?"

Shahram Anver: Uh. Mm. "Pretty sure it's okay, but, you know, let me know." Um, there'll probably be some conventions that the SRE has written, like, here's how you do incidents, here's how you investigate.

Demetrios: Yeah, I was gonna say that, that feels very synchronous, and there's probably a world where it's- Happening. -happening in parallel or it's-- they're happening together or- Yes.

Demetrios: Yeah, like conventions might be a very rudimentary way of looking at that, or they're just-- o-one is writing the code, and the next thing you know, there's not even like a PR that needs to happen because one writes the code, the other one comes and refactors it- Sure. Sure ... as it's happening, right?

Shahram Anver: Sure. But I think, like, at least on the human side, it's more the recognition that you have these two entities who are experts in their own domain.

Demetrios: Mm.

Shahram Anver: And we've found very slow ways for them to talk to each other. But, like, we've-- we agree that these are two separate entities because they have very separate objectives, so they, they do belong in the org. Yeah. But, [00:58:00] um, you know, they've got to talk. So I don't-- I just think that, you know, by twenty twenty-seven we definitely will have figured this out.

Shahram Anver: Like, I know, like, we are thinking about this deeply. Like, how do we integrate even better with, uh, you know, the Cloud Codes and the CodeKs of the world- Sure ... and how do we actually bring back all the knowledge that we've gained from production so that, you know, uh, you're giving it to the coding agent, like-

Demetrios: So the coding agent doesn't make the same mistake.

Shahram Anver: Exactly.

Demetrios: Yeah.

Shahram Anver: That to me is value, right? Like, that's, you know-- I-- Like, one of the things which is, like, my big pet peeve is You know, the whole point of SRE is you're trying to protect the reliability for your customers. It's not to reduce incidents. Like, that's a byproduct, right? Like, the whole point is, like, that's your main goal.

Shahram Anver: So if you can prevent the incident, that's what you should be doing, right? So say like any model where it's like, oh, I've improved MTTR and I've-- like I'm on jumping on every incident- My F1 score ... like, who gives a [00:59:00] fuck?

Demetrios: Like- It's just like back in the day with the machine learning models and being like, "Wow, my F1 score is-" Right

Demetrios: amazing. Right. And then you're like, actually, nobody uses that model. Right. So good job. Like way

Shahram Anver: to miss the point. Yeah. Exactly. Exactly.

Demetrios: Oh, man. Uh, I did wanna s- ask you about how you're bringing the learnings back from-

Shahram Anver: Hmm ...

Demetrios: production into the coding agents.

Shahram Anver: So I, I mean, I, I think it's rudimentary right now.

Shahram Anver: So, um, today you can have the coding agent speak to Cleric through MCP, um, and, you know, you kind of like get Cleric to weigh in, um, on what it's-- on what it thinks. Uh, we also have, you know, like these conventions that you can have in the repository, which we know that the coding agents also use, things like that.

Shahram Anver: Uh, but I don't think that that's the way it's gonna end. I, I think we, you know-- Like there's already kind of, um, attempts. I, I know like Google's working [01:00:00] on A2A. I'm not sure how much like the adoption has been- Mm-hmm ... but like you-- we're trying to create these open standards for these agents to, to talk to each other.

Shahram Anver: Yeah. Right? We haven't adopted any of these yet, uh, just to be upfront, but I think they will become more and more of a thing. 'Cause to your point, um, you don't want it to be synchronous. You want it to be like info.

Demetrios: Yeah. And how do you just keep it so that it's not blowing up the context when you do have whatever it is that you wanna give it as much information about what you've learned from production.

Shahram Anver: Mm-hmm.

Demetrios: But you don't necessarily wanna give it everything because then you're just blowing up the context every time. And so is it where you need to even add it to the context, or you just have that sub-agent that's coming in behind every coding agent that is your sub-agent, SRE agent that- Mm-hmm ... says, "Oh yeah, I know about these things, and I know that this happened recently, and it has all of the necessary [01:01:00] information."

Demetrios: And, like, that also doesn't feel like the right way to do it, you know? So have some cleanup agent that is- ... constantly going behind the coding agent. Then, uh, it's just like, but how do you make sure that you do get that 360 degree, you have the security, you have the, um, the stability. You have all- Yeah ... of these things that you're checking for.

Demetrios: I, I've heard when we had Sid do the coding agent conference, he said that one thing that they'll have is an adversarial agent-

Shahram Anver: Mm-hmm ...

Demetrios: that will try and just, you know, find problems with whatever it is. And so that's one way that they'll do it. But Really relaying the information from production to recognize that, "Hey, this happened.

Demetrios: It's the second time that [01:02:00] this has happened. Let's not have a third time. Right, guys? Come on . Make sure." Oh, I just-- I f- m- wanna see what the dynamics are of the agents that you wake up to, and the SRE agent is just Debbie Downer all the time. Like- ... you know, this- Don't you? ... PM agent is constantly shipping crap, and I have to clean up after him.

Demetrios: What is this?

Shahram Anver: Well, you know, like it-- that's happening now, and it's already just-- it's, it's hilarious. Uh, like we, we have a sub-agent that we call the investigation auditor.

Demetrios: Mm-hmm.

Shahram Anver: Uh, and the investigation auditor's job is to be Debbie Downer. Mm. It's like, "You have not done a good enough job. Keep going." And here's, like, why you screwed up and things like that.

Shahram Anver: Then, you know, sometimes the main agent will complain saying, like, "This is, like, too many times now. Um, I, I think I've-" Give up? "... done the best I could. So I'm just gonna return back to the [01:03:00] user."

Demetrios: No

Shahram Anver: way. So yeah, it's, it's, it's really funny just reading through these traces. Um, but it's funny you mention that because I think, you know, like the stuff I talked about at the beginning, like, you know, the ant colony and things like that, like, I think we've got to the point where we're now doing a lot of experimentation on what are the right abstractions, right?

Shahram Anver: Mm. Because I think, you know, for a long time it's been sort of the main agent only, then it was main agent, sub-agent. Like, I think Anthropic started releasing, I think it's still in beta, where you actually have, like, agent teams- Yeah, agent teams ... which has their totally different agents, and they don't share any context, and they're just, like, talking to each other through, you know, some form of protocol.

Shahram Anver: So I think we'll just have more and more of, um, these things. Mm-hmm. I think, uh, I think to me, like the coding agent, SRE agent, the team of agents paradigm makes much more sense because, you know, coding agent doesn't need to know the internals of the SRE agent. The coding agent just needs to know that, hey, that's his expert that I need to talk to [01:04:00] about X.

Shahram Anver: Listen to. Yeah. And, you know... Uh, but, you know, I'm not, I'm not gonna totally listen to this agent. I just wanna get some, you know, like... Isn't that like a kind of a consulting thing where it's like the RACI matrix, like responsible, accountable, um, consulted, informed?

Demetrios: Yeah.

Shahram Anver: I think you'll have that for agents, right?

Shahram Anver: Mm. Like, you know, you must listen to this SRE agent. Oh, no, you m- you may consult the SRE agent and things like that. So I think these kind of like governance structures that we come up with are gonna be very, very valuable.

Demetrios: Yeah. A lot of the software that we use and love is going local first.

Shahram Anver: Mm.

Demetrios: So as opposed to me building something and exposing it to the world, I build something that I expose to myself.

Shahram Anver: Mm.

Demetrios: And so because of that, I don't necessarily need to have all the security and all of the-

Shahram Anver: Sure ...

Demetrios: it doesn't need to be battle hardened for production if it's just me as the main consumer, 'cause it's very bespoke software.

Shahram Anver: Mm-hmm.

Demetrios: [01:05:00] And now even more so I'm looking at it like, wow, a lot of things that I used to build, I don't even need to build anymore because I just create them as a skill.

Shahram Anver: Sure. Yeah.

Demetrios: And so it doesn't require a full-fledged app.

Shahram Anver: Mm-hmm.

Demetrios: It just requires a very clean skill.

Shahram Anver: Right.

Demetrios: And so the f- f- the act of moving from production grade software to skill, which is very, very thin and it's like the lightest form of a application possible. Right. Right? Right. It's just marked down.

Shahram Anver: Right.

Demetrios: So that's, that's kinda how I look at it. Not to say that- The software is not. But like, I still feel like there's a lot of very valuable software out there.

Shahram Anver: Mm-hmm.

Demetrios: And so I'm not fully bought into the software's cooked thing.

Shahram Anver: There's just a class of software that you're just [01:06:00] never gonna wipe code, um, in my view.

Demetrios: Um- Like payments processing, Stripe, you're

Shahram Anver: not gonna- That, that, that's a good one. But like, you know, some-something like Say, like BigQuery, for example, right? Like we dump a lot of our data in there just for analysis, things like that. I'm not gonna web code that. Mm. Like, I mean, it's just-- it doesn't, doesn't make any sense.

Shahram Anver: Like the-- this, this is a product that probably took years and years and years to get right. Um, and there's almost like a level of accuracy that you expect. Like when I write a query, I get it exactly like what I want back, and it's performant and all that stuff. So, you know, physics still applies, right?

Shahram Anver: Like there's cost of having an engineer thinking about writing the software, making sure you're maintaining it. 'Cause it, it's, it's always been true with software that the har-- easy part is building it, the hard part is maintaining it. Yeah. That's the reason we exist.

Demetrios: Yeah.

Shahram Anver: You build a whole bunch of shit, it's in production- Yeah

Shahram Anver: now you gotta maintain it. And now most of engineers are spending time maintaining [01:07:00] it and not actually getting to create. So our whole value prop is, you know, you can go create now because we can take some of that work, but the cost is real, right? Yeah. Um, but yeah, you're right. Like there's gonna be, you know, part-- some, some software where, I don't know, there's like twenty features, but you only need one.

Shahram Anver: Maybe that's just a skill. Maybe that's like a little app that you build. I think where things get more interesting to me is like in that, in that like kind of middle ground where I feel, and I'm pushing for this internally a lot, where we wanna agentify everything we do.

Demetrios: Mm-hmm.

Shahram Anver: So I mean, I, I remember like, like a year ago, um, William and I were just really frustrated by how long it took to get a website update done.

Shahram Anver: It just-- It was like, you have so many ideas, but it's like you need to-- There's a meeting, you gotta talk to this person, that person, blah, blah, blah, blah, blah. And I think like a few months ago, like all right, that's it. [01:08:00] We're moving to Astro. Yeah. Everything's gonna be code. Our whole marketing site is basically an Astro site now.

Shahram Anver: It's no longer on any kind of no-code platform. And man, it's amazing. Like it is-

Demetrios: So much faster.

Shahram Anver: So much faster. That's

Demetrios: wild.

Shahram Anver: It's-- And so, so that's gonna happen, right? Like i-it's like if your platform is not agent native-

Demetrios: Mm. -

Shahram Anver: in that agents like talking to you and working with you, then that's, that's trouble.

Demetrios: Yeah.

Shahram Anver: Right?

Demetrios: The part that is worth considering is, again, what's durable? What are the things that you're not gonna vibe code?

Shahram Anver: Sure.

Demetrios: Versus what are the things that you could probably get away with vibe coding? Or what are the things that you definitely need to be agentic first, or you want to be able to use with coding agents?

Demetrios: Yeah. You wanna be able to have that seamless experience with coding agents, and so like [01:09:00] looking at those, that spectrum is another good one. And I've-- I feel like the system of records thing-

Shahram Anver: Yeah ...

Demetrios: that's probably not going anywhere, 'cause that's pretty hard to do. Databases, like you're saying, BigQuery is another hard one.

Demetrios: Like w-what, you really are gonna go vibe code a database? Yeah. There's very performant databases, and they've been performant-

Shahram Anver: Tested through like-

Demetrios: Yeah ...

Shahram Anver: years and years and years. Yeah, I think it's almost like a dimension thing, right? Which is like, how hard is it to actually build? And what is the friction it creates for you to get an agent to work with it?

Demetrios: Uh-huh.

Shahram Anver: And then if that friction is high enough and the amount of time you need to build it is not- Oh, that's a

Demetrios: nice little diagram.

Shahram Anver: Yeah.

Demetrios: Yeah.

Shahram Anver: Now that's gonna be white coded. Yeah. I'm sorry, but it's, it's gone.

Demetrios: Yeah. But if- You're feeling that frustration- Yeah ... of like, "I really want my agent to be able to plug into this."

Shahram Anver: Yeah.

Demetrios: And, [01:10:00] "Uh, could I build it? Mm, let's try. All right. 'Cause I wanna be able to have something that my agent can plug into easily." If my agent builds it itself, then it's definitely gonna be able to plug into it.

Shahram Anver: Hundred percent. Like, I mean, I, at th- at, at this point, I mean, I've, I've, you know, I have an agent that I've connected to, like, all my stuff, like my Gmail, my, uh, you know, HubSpot and everything, and I'm constantly looking for, like, okay, how do I make this thing faster and better?

Demetrios: You know what the biggest one for me was? My buddy JQ said, "Hey, we got rid of Linear. We just went to all markdown files." Interesting. Our whole thing is all markdown files, and agents are able to do whatever with the markdown, and they update it, and they keep state on how the files are going. And so he's like, "W- we realized, like, what are we paying all this money to Linear for?

Demetrios: We don't even need it." 'Cause I don't-- He s- he was saying, "I realized I wasn't going into Linear and looking [01:11:00] at it, and all I was doing is asking the agent, 'Hey, what's the status on this Linear issue?'" And so-

Shahram Anver: Interesting ...

Demetrios: there was no need to actually have Linear. But then the funniest part, and it's full circle moment, he said, "Yeah, so, uh, now we realize there are people that like to see the status, so we're now vibe coding a front end."

Demetrios: So I'm like, "Wait a minute."

Shahram Anver: So I mean, uh, the irony would be if you, I guess that's what you're implying with going full circle- Yeah ... you're kind of building Linear

Demetrios: again. He's been i-- Yeah, in a way that is much more lightweight, and he doesn't necessarily need... Like Linear, I thought it was durable, but it, at the end of the day, it's not really durable, right?

Demetrios: Like, all of your- Interesting ... your status can be markdown. And so if everything is markdown, and then it can grab context from other markdown files. So if you're, you really go all in on a file system-

Shahram Anver: Mm ...

Demetrios: then why do you need Linear- Right ... [01:12:00] to be the middleman?

Shahram Anver: I would disagree there, but like I, I think for everybody it's different.

Shahram Anver: Like to me, in that same, uh, dimensions, Linear is very low friction for an agent to use. Mm-hmm. Like I think the-- Like, you know, we use it internally, so, uh, although I do wish they'd release a, um, like a official CLI. Mm-hmm. Um, I'm not a big fan of the MCP. But, you know, generally, you know, there's so much training data on Linear that, uh, you know, agents are pretty good at using them.

Shahram Anver: And yeah, the, the effort to white-code it is probably not that high, but since the friction is so low, you're just like, it's not really worth- We don't need it ... the time to be thinking about this thing. I'd rather go focus on something else.

Demetrios: Yeah.

Shahram Anver: Right? So like, uh, to me, it's like the friction has to be really high that it's, uh, like actively impeding me for me to think it's worth like going ahead.

Shahram Anver: Because like and to the point, like now am I building a Linear alternative or am I building my business? Yeah. [01:13:00] Like that's, that's the way I'm thinking.

Demetrios: I wonder if it is for JQ, the friction was just in him seeing the bill.

Shahram Anver: Could be. Yeah,

Demetrios: it could be a cost thing. And that was what he was like- Right. Right

Demetrios: "Man, what is this?" Like it, it just eats at him, and so maybe it was just like a- Could

Shahram Anver: be ...

Demetrios: a little bit of a side project that became a, a bigger project and he realized, "Wow. All right. Let's-

Shahram Anver: Yeah ...

Demetrios: let's just see." And if you look at the bill, I don't think it's that expensive.

Shahram Anver: Right.

Demetrios: So he really probably was like, "Ah, let's just try."

Demetrios: But he also is trying to get agents doing everything. Like he has five different names of agents in his Slack that he'll- Sweet ... call on- Right. Right ... to do things, you know- Right ... because they're the different agents that, oh, this is the whatever, Linear knockoff agent.

Shahram Anver: Right. Right. Linnee.

Demetrios: Linnee.

Shahram Anver: Yeah, like, uh, but, but actually it's [01:14:00] ironic that we're talking about Linear because I think the CEO of Linear, I don't-- I can't remember his name, Curry, I think.

Shahram Anver: He's probably got one of, like, this quote that I, I really liked. Um, he talked about how, um, like the whole point of building a company is, like, you gotta be focused on your main quest.

Demetrios: Mm-hmm.

Shahram Anver: Side quests will keep coming, but you gotta focus on your main quest. Like, that is, like, your number one thing as a, as a founder.

Shahram Anver: And the thing with the agents is that the side quests have become really tempting.

Demetrios: Yeah. '

Shahram Anver: Cause you can... You know, you look at, like, what you're... You're like, uh, just-- I'm just, like, reflecting as, you know, CEO. Yeah,

Demetrios: what Jake Yu did. He was like, "Uh-"

Shahram Anver: Again, I don't know. Like- "I'm building-" ... it could be, like, a val-- So I can't, I can't speak to, like, whether that was side quest or not.

Shahram Anver: Like it seemed- Yeah, we'll have to get him on

Demetrios: here.

Shahram Anver: Yeah. Like I- We'll see ... I don't wanna roast him on, like, something I don't know. But- ... like, one of the things which I, I've s- I've seen is that being an engineer, founder is actually one of the hardest things because, like, coding feels so [01:15:00] good.

Demetrios: Mm.

Shahram Anver: You feel like you're being productive, you're getting shit done.

Shahram Anver: And when you see stuff like this, you can be like, "Oh, I can help my team. I'll just, you know, build this thing." But then, like, to me, it's like, well, that's not what's gonna make Clerk successful. I, I don't think, you know, like, my team needs me to do like a- Mm. So the only thing that, uh, like we've really, like, wipe coded seriously, um, outside, like Clerk is, um, like this, like these one-off kind of MCPs or CLIs that makes it easier for us to maybe, like, respond back to a customer or something like that.

Shahram Anver: But all the tools that we're building is really either to improve the CI/CD or something like... But if it's does a thing, it's working, and it's not crazy expensive, it's fine. Just leave

Demetrios: it. Yeah. You know, it's, uh, the exact opposite of what my talk was in the coding agent conference- ... 'cause I-- My whole talk was called Enjoy the Side [01:16:00] Quest.

Demetrios: Right.

Shahram Anver: Right. Right.

Demetrios: And so it was saying how... I, I gave the example of basically me being like, "Oh, so I wanted to build an app, and, uh, I wanted to schedule it." So then I start learning about like, oh, GitHub workflows, and boom. So then that's cool, but then I wanted to, um, make sure that- It was using a local model because it started to get too expensive when I was-

Shahram Anver: Sure

Demetrios: doing this. So then I grabbed and I, I started learning about, like, quantizing models and running it on my own machine. And then I was like, "Ah, well, actually, you know what? I need to get a GPU." So then I start learning about hosting GPUs and this and that. And so then there's this whole side quest that I go on, and afterwards, I started out just wanting to build, like, an RSS reader, and I ended up, you know, an expert in Kubernetes type thing.

Demetrios: And so it's

Shahram Anver: like- That's [01:17:00] awesome. But that, that, that's almost like-- I think that it's almost like, you know, if you're doing your main thing, then if you recognize the side quest for what it is, that's very different to deluding yourself to saying the side quest is actually helping your main quest.

Demetrios: Uh.

Shahram Anver: Right? Like, I vibe code stuff on the weekend. Like, I've, you know, like the agent, like I-- 'cause a big part of it, I mean, I tell myself it's because I wanna be in touch with the agent and stuff, but like, who am I kidding? Like, I enjoy it. Like, I, I like it. But, you know, I wanna-- but I, I just don't want to tell myself that, "Oh yeah, I did all this stuff," and like, that's definitely pushing Turric code.

Shahram Anver: Like, it's clearly not, right? It's- It's clearly not ...

Demetrios: it's just a hobby. It's another- Yeah ... hobby.

Shahram Anver: And hobbies are awesome. Yeah. Like, you learn so much.

Demetrios: Yeah.

Shahram Anver: Like, right? That's true. Like, I-- you-- and you never know what you don't know. You, you're trying new stuff, and then you learn something.

Demetrios: I'll tell you what, man, there is no better feeling, no bigger dopamine hit-

Shahram Anver: Yeah

Demetrios: when people talk about [01:18:00] social media giving you dopamine. Vibe coding- Oh, I

Shahram Anver: saw that.

Demetrios: That

Shahram Anver: was your post.

Demetrios: And, and- Yes ... and getting an open source and basically vibe coding a bug fix and then getting your PR merged to main, that's some dopamine right there. That is just

Shahram Anver: Um,

Demetrios: that's our new slot machine. That is what we've got

Shahram Anver: It, I think that's part of, like, why it's so addictive, right?

Shahram Anver: Because it feels so productive.

Demetrios: Yeah.

Shahram Anver: And there's nothing... Like I, I, I've always, like, this is, like, total tangent, but, like, some of the more popular YouTube videos are, you know, people who go and, like, mow the lawn.

Demetrios: Mm.

Shahram Anver: You know? Like, people ha- like, you just have this need to just see something get fixed.

Shahram Anver: Oddly

Demetrios: satisfying, yeah

Shahram Anver: Like something gets cleaned up or something. Dude,

Demetrios: I actually got into really deep cleaning of my car-

Shahram Anver: Right ...

Demetrios: because of too many damn TikTok videos. And so I would break out a toothbrush and get in [01:19:00] the all of the cracks- Oh, man ... and crevices and- And did you take

Shahram Anver: a before and after?

Demetrios: No, I w- I wouldn't film myself.

Shahram Anver: Right.

Demetrios: But I just, uh, and the oddly satisfying part of watching someone do it, I just got, I got influenced, you know? Right, it's so

Shahram Anver: good.

Demetrios: To do it myself. And so now m- uh, my hobby has become deep cleaning the car, which, uh, we make a day out of it. I bring my kids and have them work on it, too.

Shahram Anver: Yeah. Well, that

Demetrios: sounds- 'Cause they're the ones who make it dirty. They might as well help clean it up.

Shahram Anver: Right. Well, it also sounds like you deep cleaned this open source repository, too- Yeah. ... with the PR, right? Like, that's the point.

Demetrios: Yeah, yeah,

Shahram Anver: yeah. You know? Like, it's like this thing of like, "Oh, I fixed it. I made it a little bit better."

Demetrios: Yeah, and so you see that, and it's oddly satisfying. Yeah. And then you get that dopamine hit, and you go, "Mm, I helped. I did my part."