# Build fences not sandboxes, earn the currency of trust, and share the cognitive burden of agents across engineers | Dev Interrupted Powered by LinearB

> This week on the Friday Deploy, Ben and Andrew evaluate the arrival of Ox Alpha and unpack the security risks surrounding black-box AI tools. They examine why autonomous agents require structured fences over restrictive sandboxes, how multi-agent graphs power modern software factories, and tactical steps for engineers driving team-wide AI adoption.

_This is a markdown rendering of a live HTML page on linearb.io, generated for AI/LLM consumption — it is not a markdown-only site. To get the full HTML page instead, request this URL with an explicit `Accept: text/html` header (no wildcard, no markdown preference)._


```json
{
  "@context": "https://schema.org",
  "@type": "PodcastEpisode",
  "name": "Build fences not sandboxes, earn the currency of trust, and share the cognitive burden of agents across engineers",
  "description": "This week on the Friday Deploy, Ben and Andrew evaluate the arrival of Ox Alpha and unpack the security risks surrounding black-box AI tools. They examine why autonomous agents require structured fences over restrictive sandboxes, how multi-agent graphs power modern software factories, and tactical steps for engineers driving team-wide AI adoption.",
  "url": "https://linearb.io/dev-interrupted/podcast/ai-agent-fences-ox-alpha-trust-software-factories",
  "datePublished": "2026-08-28T12:00:00.000Z",
  "partOfSeries": {
    "@type": "PodcastSeries",
    "name": "Dev Interrupted",
    "url": "https://linearb.io/dev-interrupted/podcasts"
  },
  "actor": {
    "@type": "Person",
    "name": "Andrew Zigler",
    "jobTitle": "GTM Engineer",
    "worksFor": {
      "@type": "Organization",
      "name": "LinearB"
    }
  }
}
```

```json
{
  "@context": "https://schema.org",
  "@type": "BreadcrumbList",
  "itemListElement": [
    {
      "@type": "ListItem",
      "position": 1,
      "name": "Home",
      "item": "https://linearb.io/"
    },
    {
      "@type": "ListItem",
      "position": 2,
      "name": "Dev Interrupted - Podcasts",
      "item": "https://linearb.io/dev-interrupted/podcasts"
    },
    {
      "@type": "ListItem",
      "position": 3,
      "name": "Build fences not sandboxes, earn the currency of trust, and share the cognitive burden of agents across engineers",
      "item": "https://linearb.io/dev-interrupted/podcast/ai-agent-fences-ox-alpha-trust-software-factories"
    }
  ]
}
```

[Home](https://linearb.io/)

/

[Podcast](https://linearb.io/dev-interrupted/podcasts)

/

Build fences not sandboxes, earn the currency of trust, and share the cognitive burden of agents across engineers

# Build fences not sandboxes, earn the currency of trust, and share the cognitive burden of agents across engineers

By Andrew Zigler

|

August 28, 2026

![building_trust_with_ai_agents_cebc111498](https://assets.linearb.io/image/upload/c_limit,w_3840/f_auto/q_auto/v1/building_trust_with_ai_agents_cebc111498?_a=BAVMn6ID0)

This week on the Friday Deploy, Ben and Andrew evaluate the strange arrival of Ox Alpha and debate the security risks of black-box AI tools. They also explore why engineering trust is more critical than ever, revisit the controversial lines of code metric in the age of AI generation, and explain why autonomous agents need structured fences rather than restrictive sandboxes. Finally, they break down the multi-agent graph architectures powering modern software factories and share actionable advice for reaching staff engineer status by driving cross-team AI adoption.

### Show Notes

* [Ox alpha](https://oxalpha.com/)
* [Fundamentals of Trust in Engineering](https://luminousmen.substack.com/p/fundamentals-of-trust-in-engineering)
* [Conceptual integrity and counting lines of code](https://simonw.substack.com/p/conceptual-integrity-and-counting)
* [Graph Engineering: The Complete Guide to Building Multi-Agent AI Systems](https://lunarresearcher.substack.com/p/graph-engineering-the-complete-guide)
* [Fences, not Sandboxes](https://yegge.ai/essays/fences-not-sandboxes/)
* [How to Grow From Senior to Staff Engineer in the AI Era](https://newsletter.eng-leadership.com/p/how-to-grow-from-senior-to-staff)

### Transcript 

_(Disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)_

\[00:00:00\] **Ben Lloyd Pearson:** All right, Andrew. Well, I know you're feeling pumped. We just got out of this workshop that we had today, or a roundtable discussion, excuse me. We invited some friends of the show. We had Aloke Desai from, from Warp. Uh, we had Dex Horthy from HumanLayer. We had our own Dan Lines from LinearB talking about the software factory, and it was a really amazing discussion.

\[00:00:21\] **Ben Lloyd Pearson:** Like, I was actually really blown away at just, like, the engagement we got from the audience. Like, it was so much fun. Uh, which is why it's so g- great to, to attend these events when they're live because you kinda miss out on that if you're catching it after the fact. But Andrew, you know, I wanna, I wanna ask you before we kick off today's episode,

\[00:00:37\] **Andrew Zigler:** Mm-hmm.

\[00:00:38\] **Ben Lloyd Pearson:** What, what emoji represents how you think that that workshop went today?

\[00:00:44\] **Andrew Zigler:** Oh, defin- oh, how the workshop went today? Mm.

\[00:00:47\] **Ben Lloyd Pearson:** Yeah

\[00:00:48\] **Andrew Zigler:** Well, you know that there is a factory emoji, right? I've been making it my mission lately to remind folks that there's a factory emoji. There are, in fact, many cool emojis that I think are way underused, and if you follow me on LinkedIn, you might know I \[00:01:00\] often rant about this.

\[00:01:01\] **Andrew Zigler:** But, the energy in that chat was really, was so much fun. Many of them who are from the Dev Interrupted community, of course, I've just been really excited to be plugged in on this software factory story for the last month, month and a half with y'all. And if you've been following and reading along, you know I've been writing articles covering some forefront leaders on the topic and, you know, we just brought them together for this roundtable discussion.

\[00:01:23\] **Andrew Zigler:** And really what I was digging at more was less about, like, the nuts and bolts of how the factory happens. Like, that's interesting, but I was more intrigued by how do we answer for its impacts within an organization? How do we express its value to others who aren't in the nuts and bolts? And I think that's, like, a lot of work ahead of us still to do, but thankfully we had some really smart folks there, uh, to kind of show us what they're thinking as, you know, front runners on this kind of, uh, SDLC transformation

\[00:01:55\] **Ben Lloyd Pearson:** Yeah. Well, you know, uh, you know what I think about the factory emoji, to your \[00:02:00\] point is, you know, maybe we can start a movement where that starts to become the emoji you use when you need to flag things to your factory. Like anything that needs to be pulled in, you just, you just throw a factory emoji on it, and it just gets pulled right in and ingested.

\[00:02:13\] **Andrew Zigler:** Now you have to be very careful about emoji code 'cause I have lots of like, uh, like emoji code way of talking to my agents and ways of flagging stuff. So that might already be in the works. So if you're not leveraging the factory em- emoji today, uh, this is your reminder to do so.

\[00:02:30\] **Ben Lloyd Pearson:** Yeah. And these are the types of topics that we talk about here at the Friday Deploy brought to you by LinearB. I'm your host, Ben Lloyd Pearson.

\[00:02:39\] **Andrew Zigler:** And I'm your host, Andrew Zigler.

\[00:02:40\] **Ben Lloyd Pearson:** And this week we are covering a new reasoning first model, the currency of engineering trust, counting lines of code again, building multi-agent graphs, fences over sandbox, and the journey of going from a senior to a staff engineer in the AI era. So lots to cover today.

\[00:02:59\] **Ben Lloyd Pearson:** So let's just \[00:03:00\] get right into it at the top, Andrew, and talk about this new model that just launched. Um, what do you have to tell me about Ox Alpha?

\[00:03:07\] **Andrew Zigler:** Yeah, so there's this mysterious new model called Ox, uh, Ox Alpha that's available on Open Router, and what makes this one really striking not, it's not only that it's a reasoning, uh, model, has a one million token capacity, uh, in its context window, uh, but it's released under stealth, so we don't know who e- exactly is releasing it.

\[00:03:27\] **Andrew Zigler:** Uh, they were asked, they asked to be kept u- unnamed when they, when they put it on Open Router. Uh, but also the inference is zero. Yes, completely free, as of right now at least, and I could imagine this is probably a very short-term experiment. Uh, Ox Alpha's not the first of a mystery model named after an animal that was later unveiled to be maybe like a new model from an, open source lab, uh, or a foreign model as well, and, uh, it's interesting to really observe this pattern of releasing a model this way, and I've been trying to pay more attention to \[00:04:00\] model releases, especially as I've started self-hosting a lot of them myself.

\[00:04:03\] **Andrew Zigler:** And, uh, what I'm paying attention to in this one is, like, I wonder if this would be a common practice in the future. I don't know about you, Ben, but, like, I have a lot of reservations about throwing any kind of token into a complete black box like this, and I already know that I'm surrendering so much information already to the frontier providers I use, and I don't even like that as much.

\[00:04:23\] **Andrew Zigler:** But the idea of not even knowing where it's going, and then the fact that it's free just makes me wonder about what are the economics on that and where is that data ultimately going? What do you, what do you think when you see this kind of business model?

\[00:04:35\] **Ben Lloyd Pearson:** Yeah. Well, I mean, first of all, just because AI's here doesn't mean we just gotta like throw all of our security practices out the window. Uh, in fact, it's probably more important than ever. So, I mean, you gotta be really conscious about w- you know, like you said, where, where you're sending your data when you're using these models.

\[00:04:53\] **Ben Lloyd Pearson:** So, you know, I, I, I did play around with it a little bit, you know, nothing that would like really, put me at \[00:05:00\] risk of anything. Uh, just to see, you know, I just wanted to see how it sort of behaved and performed. I was really surprised actually at its speed. Like it was remarkably fast at producing outputs.

\[00:05:09\] **Ben Lloyd Pearson:** I was not super impressed with the quality of it, of what it was giving me, and I was just asking it to do like some basic summaries of general topics that I know a lot about and I can like really quickly validate, you

\[00:05:20\] **Andrew Zigler:** Yeah.

\[00:05:20\] **Ben Lloyd Pearson:** of quality.

\[00:05:21\] **Andrew Zigler:** Yeah

\[00:05:22\] **Ben Lloyd Pearson:** but, you know, more competition, cool, but these black box models that just take your data in and, you know, clearly they're using that to make money somewhere else.

\[00:05:32\] **Ben Lloyd Pearson:** So quite... You have to ask yourself like, if the platform is free, then you are the product or something like that, you

\[00:05:38\] **Andrew Zigler:** Yes. Yes, exactly.

\[00:05:40\] **Ben Lloyd Pearson:** Yeah.

\[00:05:41\] **Andrew Zigler:** So one thing to keep in mind about this is one way that you can at least stay plugged in the know about the capabilities of these models, like Ben, you talked about, like, just throwing some prompts at it to experiment or whatever, is

\[00:05:53\] **Ben Lloyd Pearson:** Yeah

\[00:05:54\] **Andrew Zigler:** idea of just having some back pocket evals. Like, I've seen a lot of really smart folks talk about this.

\[00:05:58\] **Andrew Zigler:** Like, your ... Think of your \[00:06:00\] hardest, like, non-domain specific or non-PIIE kind of problem, or think of that problem and abstract it into something that just, you know, is, is, is a more pure representation of what you're trying to achieve, and use that to test new models. And I think that if anything, this is just a reminder that it's good to have those because you can throw these at this, see how it performs, and then instantly get a sense of how it matches up against other models you use without having to actually start funneling your information into it.

\[00:06:33\] **Andrew Zigler:** Uh, so definitely a practice to pick up.

\[00:06:35\] **Ben Lloyd Pearson:** Yeah. All right. Let's move on to talking about trust in engineering. So this is an article that I, I personally really enjoyed, largely because it's, it's a simple concept, and it feels obvious, but I feel like it's something that partic- like today i- in this moment that we're at with AI taking over our SDLC, it's really important to think about the concept of trust. , a-and it seems obvious, but I, \[00:07:00\] I really wanna revisit as we're talking about this agentic software factory more and more. Think about, like, the code review. Like, in the past, you know, when you would review, like, your peer's code, it, like, rarely involved actually looking at every line of code, that they submitted.

\[00:07:15\] **Ben Lloyd Pearson:** You know, you would typically focus on, like, one or two components of it, and then hopefully, you know, you can trust the developers that you work with, and the person who wrote it enough that you don't have to look over everything with a microscope. And of course, you know, there's some exceptions to the, this rule, like when you have new people on the team or, new code bases, et-et cetera. But that same trust, it needs to extend into every single system that supports your SDLC. So this,

\[00:07:42\] **Andrew Zigler:** Mm-hmm.

\[00:07:42\] **Ben Lloyd Pearson:** is things like your CI/CD system, the security analysis that you do, um, the data that you use to inform your decision-making. Um, this list kind of goes on and on. uh, you know, because of how trust can be broken in a single moment, it's really kind of more important than ever \[00:08:00\] to, like, focus on continuously building in it, uh, building it as AI takes over more of the SDLC.

\[00:08:05\] **Ben Lloyd Pearson:** Because if you don't trust your CI/CD workflows, how can you trust AI to safely deploy software on your behalf, for example? You know, if you don't trust the quality of code that's being produced by AI, you'll want to analyze every single line of code, which is a completely exhausting experience that you really can't expect someone to do in this day and age. To successfully transition, uh, your organization into this agentic way of working, you really need to build that trust around every single aspect of your SDLC. Um, and anywhere that you don't have that trust, like, that's, that's where you need to be investing your, your human effort to resolving. Um, and that's, you know, just gonna be key to success with AI systems. What, what'd you think of this article, Andrew?

\[00:08:48\] **Andrew Zigler:** I think it's a really great summary, and trust is absolutely a currency, as we all know and are familiar with. And in this article, there's a really, a really visually s- uh, \[00:09:00\] strong graph that shows how you build trust incrementally over time. And then you might have one incident or one problem or one stumble, and all of that trust is gone.

\[00:09:10\] **Andrew Zigler:** Now you're negative trust, and it's not like give a little, take a little. It's earn a little, and then a lot gets taken if something is just, uh, breach that trust. And the thing about trust is that it goes hand-in-hand with, with some other things that, again, relate back to the software factory, like what you did, and one of them is ownership.

\[00:09:31\] **Andrew Zigler:** We talked about that a lot on, on the workshop today, about how, you know, everyone on the panel and, and in the audience too emphatically agreed that, you know, engineers are still responsible for the code. They still own the code. And just because you're abstracted from what was a traditional PR process doesn't mean that you aren't owning what's getting shipped.

\[00:09:52\] **Andrew Zigler:** And so that ownership also comes with a r- responsibility to understand. And so by owning \[00:10:00\] and understanding, this is how you build trust. And in the, a factory kind of mindset, you're really at risk of compromising one of the two or both to achieve the factory, and you're going to lose the trust that makes that possible.

\[00:10:15\] **Andrew Zigler:** So I think that this is a really strong insight. It's definitely, like, one of those things where once you start slicing it, it's like, yes, trust is so important, but this breaks down, like, the actual currency of how you build that in an agentic time. Uh, and it ultimately comes down to just having systems where humans are paying attention to what's going on.

\[00:10:35\] **Ben Lloyd Pearson:** Now let's talk about a metric that always seems to come up time and time again, and that is lines of code.

\[00:10:42\] **Andrew Zigler:** Oh no, here we go again.

\[00:10:45\] **Ben Lloyd Pearson:** So we've got an article from Simon Willison, uh, where he actually makes the argument that measuring the lines of code that are produced per day is actually a good metric to measure, um, particularly as you have more and more coding agents, deployed in your \[00:11:00\] organization. You know, I feel like it's, it's such an old topic that, uh, it, it sort of predates even, like, I, I don't know. It, it's like it's come around so many times.

\[00:11:10\] **Andrew Zigler:** It's really just like the ancient ancestor of tokenmaxxing. Like, they're in some genealogy tree together.

\[00:11:17\] **Ben Lloyd Pearson:** Yeah, so it's only natural we've come back to this now, I suppose. and I also sort of bucket like metrics like PRs per engineer or like commits per engineer in sort of a similar thread to lines of code. It just represents the volume of work really. I, I'm actually gonna take a, a very nuanced take, I suppose. Um, I actually think that, uh, both the people who say it's a bad metric and the people who say it's a good metric are both correct, and I can explain.

\[00:11:45\] **Andrew Zigler:** Okay. Okay

\[00:11:46\] **Ben Lloyd Pearson:** yeah, so, you know, these types of metrics, like lines of code, it's bad when you use them like as the goal. Like if your goal is just to generate more lines of code, it's very easy to come up with ways to do that. Um, so if you're setting like \[00:12:00\] performance evaluations around it, like, you know, that number will absolutely climb, uh, pretty much guaranteed, you know. And, and, and I know this from personal experience. I was actually on a team that just like we were measured on that, and we just wrote a script that like gave us this nice, convenient up and to the right chart that made us always look good and predictable. So in those situations, they're bad. But, but they can be good when you use them as signals for, to see that something has changed. So if you see the same engineer suddenly quadruple the lines of code that they're generating, which is happening to a lot of engineers right now, or maybe they're doubling their pull request output, this is a significant thing that you actually should spend the time to understand. Because if they're doing that while also maintaining quality, like if quality signals are still looking good and things are getting shipped into production, it's probably a net benefit that they are now producing more of what they were doing before. You know, in fact, you know, we've covered a little bit of this in our recent workshop, uh, on the \[00:13:00\] emerging productivity gap.

\[00:13:01\] **Ben Lloyd Pearson:** So, you know, if, if you haven't checked that out, if you're listening to this and haven't checked that out yet, you definitely should because there's some really incredible data in, in this. but to summarize, you know, we saw in our own customer data that since the start of this year, and really starting, you know, sort of at the end of last year, but taking off in January and February of this year, we've seen the developers who are using AI the most, so this is the, the cohort of people who are using, you know, AI at least three-quarters of their coding days. they have more than doubled the, the amount of pull requests that they are outputting over the last six months. and this correlation declines based on how much or little AI they use. So if you're someone who, who has no measurable AI activity, we're actually seeing that you probably are still shipping about the same velocity as you were a year ago. and the people in the middle, you know, are sort of in between that, you know, flat and two x, uh, ch- uh, change. so you know, n-now what... And whether or not this like means that these teams are delivering more value to their \[00:14:00\] customers or to the business, you know, that's like a totally separate discussion that is absolutely valid It's totally worth happening, having.

\[00:14:06\] **Ben Lloyd Pearson:** Um, but if you're purely looking through the lens of velocity, um, I think right now, right, just right in this moment, lines of code generated does have some validity as a thing to monitor. But what do, what do you think, Andrew?

\[00:14:20\] **Ben Lloyd Pearson:** Did I convince you?

\[00:14:21\] **Andrew Zigler:** You did convince me, so I usually do, Ben. But i- in this particular one, first off, just to take a step back, Simon Willison is nail on the head here, or had ... Right, he just hits it right on the head because, you know, uh, really what he's talking about here is the idea of conceptual integrity, understanding what's happening in your world, uh, and counting lines of code.

\[00:14:43\] **Andrew Zigler:** These... This is something that up until AI, they were very closely coupled together, uh, in terms of, like, how they moved and the capacity move. But now you're in this world where agents can produce so much more code, but, you know, your ability to comprehend and understand what \[00:15:00\] those changes are on a day-to-day basis is the same as it was, you know, back then.

\[00:15:05\] **Andrew Zigler:** And so what happens is, you lose the ability to understand what's moving in your code base, and then you lose your ability to point it in the right direction, and it kind of escapes you. And this is like a really... It's the opposite of, like, a bottleneck. It becomes like its own bottleneck, where you become divorced from what is in the reality of your product.

\[00:15:27\] **Andrew Zigler:** And teams are trying to solve this in a lot of ways. We haven't figured out how to do this. Like, observability has b- been found to be a fantastic way to gain back that cognitive understanding because you ship 10, you know, 10x more, but your capacity to understand is still that 1x it was. So you can't understand necessarily what your, your code is or what your product is by looking at the code.

\[00:15:51\] **Andrew Zigler:** You must observe it in production and get real-time feedback and signals that can go into that high productivity highway that you have with your \[00:16:00\] agents. We've explored that with, with Honeycomb and others on the show talking about that important role. And also too, it calls out the emerging problem we're having, especially now like mid, late 2026 as the models are getting much smarter and our coding loops are getting much tighter, is that we're hitting a, discernment horizon.

\[00:16:19\] **Andrew Zigler:** You know, Yegge has talked about this as well, about our ability to understand if a model is good is going to be escaping our cognitive capacity, right? And you're gonna get these models that aren't consumer grade and operate on different levels of thinking that aren't necessarily something that we can cognitively keep up with once they start going.

\[00:16:38\] **Andrew Zigler:** So in that world, like, you wanna be optimizing for how do I bring this... How do I wrangle in all of this, like, cognitiveness and compact it down so I can understand it? There's a lot of really cool tech out there for this. Like, if you haven't tried out TL Draw, this is a really fascinating way to do this.

\[00:16:57\] **Andrew Zigler:** Give your agents access to an infinite whiteboard and let them \[00:17:00\] diagram things for you. Simon Willison himself even has a very famous library called Showboat that helps your agents create, like, markdown and, uh, like Quarto n- notebooks of, like, the work that they do and, like, show it to you.

\[00:17:14\] **Andrew Zigler:** Like, imagine, like, you're a manager and like- your, your, your employees come in and, like, present their project to you. It's like that. And so there are already people tackling this cognitive compression. People just have to figure out what that looks like for their- themselves and their own team.

\[00:17:27\] **Ben Lloyd Pearson:** Now let's move on and talk about graph engineering. Andrew, what do we have on graphs?

\[00:17:33\] **Andrew Zigler:** Okay, this is a, a great kind of a comprehensive overview of graph engineering and where it currently stands, and it's the idea of multi-agent systems, whether they're running in the cloud or they're just kind of loose agents on, defined on your machine, being able to define what are their inputs and what are their outputs.

\[00:17:52\] **Andrew Zigler:** If you're an engineer or if you've tinkered with things like LangGraph, then this is the idea of thinking about the, just the agents you \[00:18:00\] use on a regular basis existing within a graph with each other, and if they were to communicate or hand things off to each other, what's the shape of the data and expectations that they would have?

\[00:18:10\] **Andrew Zigler:** And this article also gives some really good dives into graph shapes. Uh, some of these are, you're already familiar with, like where you ask three models to write you a spec and then you pick your best one or combine them into a good one, or you run three models on one prompt and you compare them and see what their differences are.

\[00:18:29\] **Andrew Zigler:** Like, these are all actual shapes, right, that can be expressed in, by agents in a graph with inputs and outputs that consolidate together. And by thinking about agents in this kind of shape, you can get really good at piping information between your sessions. Like, you know, we're gonna talk a little bit about some other stuff related to this here in a moment on our next story, but once you start cracking into this, then you get a communication language between your agents that can be stateful, like two live conversations helping each other, but also stateless, \[00:19:00\] like leaving helpful artifacts for each other.

\[00:19:03\] **Andrew Zigler:** Really smart read if you're starting to work with agents in this way.

\[00:19:06\] **Ben Lloyd Pearson:** Yeah. Uh, you know, I loved this article so much that I skipped everything and went, just went straight to the end because I was like, right from the start, I was like: Oh, yeah, obviously I need this. Like, why didn't, why didn't I think of this? Because I, I feel like I've been struggling with this exact challenge right now.

\[00:19:21\] **Ben Lloyd Pearson:** Like, in the past, you would sort of do... Most of your work was a lot more linear, you know, 'cause a human can only sort of like do one thing at a time. So, you pick up one task, you know, you plan it out, you develop it, build it, you release it, then you move on, you know, hopefully. and that's really not how work h- happens in this, this new agentic factory environment. So the reason I skipped all the way to the end is because there were a lot of really great examples of agent fodder. I got to the end, saw all the stuff they had that was like, "Here's what you give to your AI," and,

\[00:19:53\] **Andrew Zigler:** Yeah, it was a treasure trove.

\[00:19:55\] **Ben Lloyd Pearson:** Understood like what this was achieving for me,

\[00:19:57\] **Andrew Zigler:** Yep

\[00:19:58\] **Ben Lloyd Pearson:** and we're like, you know, as these \[00:20:00\] software factories are getting spun up everywhere, we're also building this like content factory over here at Dev Interrupted, which has just been fascinating to watch come together. And one of the challenges that we face is like sometimes like downstream assets actually change your, your, the stuff that you put in the upstream sources.

\[00:20:16\] **Ben Lloyd Pearson:** So, think about it like this, as you're building something, your understanding of the problem that you're solving might change along the way. Um, and when that happens, you need to change the upstream spec to match your new understanding. Um, and in content, this can happen like quite a lot actually. Like when humans were building it all, like, like I said, this was all sort of a linear thing. But now that we're constructing these factories, we have to be thinking of things more as a series of components or, you know, I've been calling them like assembly lines, I guess. but these components that you can sort of bring together and control like the inputs and outputs and, and, and have different graphs of ways that they work together. So a lot of great agent fodder i- in this is the short of it, so definitely go \[00:21:00\] check it out.

\[00:21:00\] **Ben Lloyd Pearson:** All right, Andrew. Of course, not a week can go by where we don't cover the likes of Steve Yegge, it seems like. Um, and he's back again with another really great article about why we should be thinking about fences rather than sandboxes. Uh, so of course, we've been hearing all these stories about AI escaping the sandboxes and, and doing all sorts of crazy things, and everyone's building san- AI sandboxes to control them. Yegge is out, actually out with an article now that, where he thinks that rather than thinking about containing AI in that way, turns out that a fence metaphor may actually be, um, the better solution for this. So, you know, Andrew, I know you and I, we love thinking about metaphors for our factories. They are a surprisingly effective system for, getting consistent behavior, even though it does appear odd often, the, the sort of the vernacular and the systems that, uh, they will adopt around it. But I really love fences as a concept, um, because I love the idea of having something that \[00:22:00\] outlines to agents when they're out of bound from their listed responsibilities. You know, to me, a sandbox sort of feels more like it's a trap that you should escape from. It's like, "Oh, you, you're stuck here.

\[00:22:12\] **Ben Lloyd Pearson:** You need to get out and, like, get more and do other things." Whereas a fence is something that was, like, deliberately constructed to make sure that you're gon- you're not gonna cause harm, like, either to yourself or to, like, something else that's out there. It's like somebody put this here specifically to, to put, to tell you to go the other way, you know? and then, yeah, and then not to mention, like, there's this whole story in, you know, speaking of metaphors, about how Yegge's factory has this entire legal system that's based on, like, this game that he's building that I know you love. uh, it's just, again, just such a great article from Yegge.

\[00:22:44\] **Ben Lloyd Pearson:** So what'd you think about it, Andrew?

\[00:22:46\] **Andrew Zigler:** Well, obviously a zillion things because I love reading things from Steve Yegge. It feels like it's always a, like a message from the future that's just kinda washed up on the shore whenever I read it. Um, I, I, I usually am really intrigued by some of the \[00:23:00\] ways that he gets really ahead. And

\[00:23:02\] **Ben Lloyd Pearson:** Mm-hmm.

\[00:23:03\] **Andrew Zigler:** from his practices that I've certainly adopted into my own harness.

\[00:23:06\] **Andrew Zigler:** And so what's been fascinating is because the nucleus of what I work with on a regular basis, like going back to the content factory that we have here and really my, my agents running otherwise, is that the nucleus of them really comes from these ideas from Yegge. So I've also been able to observe, interestingly, my own harness and environment evolve alongside his as, you know, he, he makes and remakes it.

\[00:23:29\] **Andrew Zigler:** And when he, when he comes with these articles and shares them and he shares something that turns out is existing or, you know, I have in mind or it exists in some other form is not only really exciting to me, but it's a great clue that we're finding the right shapes that work because, independently arriving at those kinds of conclusions is a, is a great supporting evidence for why they're effective there in the first place.

\[00:23:55\] **Andrew Zigler:** And fences is the latest from him, and it's something that I've had a lot of luck with as well. \[00:24:00\] And all fences are, are just textual reminders and notifications and otherwise text that gets into the context window when an agent starts to go down a certain road. And I, I loved your, um, the way you talked about the sandbox and how it kind of induces a panic mode.

\[00:24:16\] **Andrew Zigler:** I've definitely experienced that. There are definitely some agents that I use, especially ones that I have just kind of read things on the open web, that I purposefully sandbox for those reasons, right? And I notice that sometimes when they have issues or if they encounter a problem that I need to improve within, like, their bubble wrap, they kind of, like, panic a little bit and start doing all sorts of random things just, like, trying to solve it.

\[00:24:39\] **Andrew Zigler:** And so, a, a fence is a much more subtle reminder. It, it's a nudge that like, "Hey, maybe this isn't the right thing to do," or, "Hey, maybe you should check with so-and-so." And that is therein the really fascinating stuff that starts to emerge, is because then you actually have agents and realms of responsibility that are in charge of whether or not agents can make those \[00:25:00\] decisions.

\[00:25:00\] **Andrew Zigler:** And it all sounds arbitrary, but agents are incredibly effective actually at building shapes like this. If you, uh, bring to it the problems that you're trying to solve and y- you explain to it in the capacity that Yegge does, where you have individual, like, seats or identities for your agents, what does emerge pretty quickly, and what he's commenting on throughout this article, is more like a governance system, like laws and rulings, rather than a mechanical thing, rather than, like, actual, like, pipes and stuff.

\[00:25:34\] **Andrew Zigler:** Like, he- his own reaction was verbatim in his article. Like, he opened it up and he saw government instead of tooling, and he saw constitutions, courts, jurisdictions, case law, rulings, and registries. He said, quote, "Straight up WTF, no words." And that's what they built, and those are fences. Those are things that are gonna pop up in context windows that nudge agents off of certain things, and I definitely \[00:26:00\] have observed agents kind of falling into this kind of formation.

\[00:26:03\] **Andrew Zigler:** Really fascinating, uh, exploration from him. But I'm-- I wanna challenge Steve on, on one thing here actually, because in my position of having, you know, created and worked with my own harness as well and having a lot of these parallels, he dives into how they, you know, really are hungry to make this kinda government.

\[00:26:21\] **Andrew Zigler:** Well, I kinda just feel like it's almost like inviting a Lord of the Flies situation, because really what he also comments on is how they have the reasoning and judgment of sixth graders, even at best. So when you put them in charge of creating what kinda ruling government system they think is gonna be effective, it kinda becomes like a caricature of itself.

\[00:26:43\] **Andrew Zigler:** And Yegge's even made these articles recently that makes fun of how his agents talk to him because they're so cryptic in how they explain this vernacular. So you have to fight that. Um, that's definitely something I've been fighting. And, uh, the last note that I'll leave it on is he comments in \[00:27:00\] here that software factories are real.

\[00:27:02\] **Andrew Zigler:** He calls Wheelhouse a software factory because it is. and because he alludes to them being more like school kids, it's more like, I think, like a classroom is the way that these things need to be run. And so by providing these opportunities for them to create structured ways of enforcing rules and when things are due and who's in charge of what, like think of like a classroom where someone's in charge of turning off the lights or cleaning up after whatever.

\[00:27:29\] **Andrew Zigler:** Like that's the kind of rule system that actually cognitively works better for them I find. Uh, really fascinating dive. Um, maybe we'll get another one from him next week and I'd love to nerd out about it again then.

\[00:27:42\] **Ben Lloyd Pearson:** Yeah. You know, you have me thinking, your comments, ' the sixth grade reasoning really did stand out to me as well because it, it really made me, like, imagine what... At the scale that he's operating at with his, his AI agents, it's like, i- imagine if that was real life and you had an army of, like, 50 sixth, sixth graders.

\[00:27:59\] **Andrew Zigler:** \[00:28:00\] He's, he's a counselor at, at, at, at summer camp, right? Imagine like the coolest counselor at summer camp, and that's Steve Yegge and his army of wheelhouse agents that are costing him

\[00:28:10\] **Ben Lloyd Pearson:** Yeah

\[00:28:10\] **Andrew Zigler:** a, how much a month is he at now? $120,000 a month equivalent in API tokens, and he's paying 5K out of, out of pocket.

\[00:28:19\] **Andrew Zigler:** So

\[00:28:19\] **Ben Lloyd Pearson:** But it, but it's, and it's been amazing how we've, we've been able to track the pr- the progress of these models sort of like, uh, because I think two years ago I was sort of saying that these models, th- they reason at about the level of a three-year-old, you know?

\[00:28:31\] **Andrew Zigler:** Mm-hmm.

\[00:28:32\] **Ben Lloyd Pearson:** like they're very smart, like the most intelligent three-year-old you've ever met, you know. Um, and now to see that we've kind of stepped up to like the sixth grade level, like that's a pretty significant step. But it does make me like wonder like what does it look like when it's in like an army of high schoolers and they're like rebellious and like you know, like, like way more like irrational and like capable at the same time, you know.

\[00:28:55\] **Ben Lloyd Pearson:** Is that what's next?

\[00:28:56\] **Andrew Zigler:** I think, I think we're gonna find out.

\[00:28:58\] **Ben Lloyd Pearson:** Yeah. \[00:29:00\] Uh, but, you know, and, and to keep things lighthearted, there's also just a really great comic that he added to this about like, passive comments that he makes into his terminal, um, become these, like, decrees that, like, kick off multiple days of work and how it's... Yeah.

\[00:29:17\] **Andrew Zigler:** that's so relatable too. I'm like, "Guys, what did you do?"

\[00:29:21\] **Ben Lloyd Pearson:** Yeah, yeah. So yeah, yet again, another great article from Yegge. All

\[00:29:26\] **Andrew Zigler:** Mm-hmm.

\[00:29:26\] **Ben Lloyd Pearson:** right, but let's move on to, um, how to grow from senior to staff engineer in the AI era. Uh, so this is a really great article from a staff engineer over at Pinterest that sort of lays out the dimensions that separate a staff engineer from like a senior engineer. , the main themes of this were, you know, you wanna expand your surface area beyond your job description. You want to influence teams that don't have any sort of accountability or responsibility to and then also, like, to be able to build scalable systems that continue to run \[00:30:00\] without you. and, you know, I think really the core message of this article is that, you know, if you wanna have that staff-level impact, it really comes from work that makes other people more effective, rather than any sort of, like, individual work that you do on your own. There was a line in here that stood out to me quite a bit, where the author said, "Nobody at Pinterest asked them to become the AI person."

\[00:30:21\] **Ben Lloyd Pearson:** Like, no one said that they should go and be the AI person. They saw that this was the next problem that their organization was going to need to solve, and they got ahead of understanding how to reach, you know, how to, how to be the, the solution there. And I think that's such a succinct point on, like, what it really takes to reach, like, the top tier of IC leadership roles. This goes back to a practice that I remember learning, uh, back when I was in, in grad school, because one of our teachers back then really enforced this idea that, like, anyone can poke holes in something, um, and tell you what needs to be \[00:31:00\] fixed. But leaders are the ones that provide the solution along with those critiques. And the reasoning to justify that solution as well.

\[00:31:08\] **Ben Lloyd Pearson:** So, know, for example, it's like instead of leading with the negative and telling somebody that, that something they've done is wrong, um, you actually wanna show up and provide the solutions that, you know, helps them get better along with you. Lots of really great advice in this article about measurement and reporting which is really important, uh, particularly if you're at a larger organization. Um, so yeah, a lot of, a lot of great advice on just what it takes to be at that staff level. So what did you think, Andrew?

\[00:31:36\] **Andrew Zigler:** The cross-team enablement is the real standout insight here. For me at least, that's definitely something that I'll, I'll carry with me because I think that is what really stands out for, like, what you called out, like, a top of their career kind of IC. The IC that is so good at their job that they're, like, prescient.

\[00:31:53\] **Andrew Zigler:** They know what you and the company and your team is gonna need, and they've thought about the alternatives, and they've \[00:32:00\] ideated it, and they've delivered at least a bridge. Maybe it's even like you're literally walking across, like, whatever they've laid out just to get to the other side so that it's not an obstacle.

\[00:32:10\] **Andrew Zigler:** And those kinds of ICs that help create and open the path ahead so people can get moving, it requires a lot of, like, a forward thinking, but also cross-team collaboration. And oftentimes, all of that work is to fix a gap or a bottleneck or an issue or something that might not even live very near to the world that's their day-to-day or matter specifically on their deliverables.

\[00:32:35\] **Andrew Zigler:** It's about thinking about what is going to enable the broader team to move towards that goal, and that's what, that's what really makes these kinds of, staff, leaders kind of stand out amongst their peers, I think. So, really cool. There were, there were some specific insights in here, like what you alluded to around how to give, like, actionable feedback to get people on board with your cross-team strategy.

\[00:32:54\] **Andrew Zigler:** But also just about, like, owning an idea or being synonymous with a \[00:33:00\] word or a concept or an innovation at your company is one of the best investments that you can make in yourself. And what I, what I mean by that is he saw an opportunity of we're gonna need to discuss as a team cross-collaboratively in one unified place about how we're experimenting with and sharing AI.

\[00:33:18\] **Andrew Zigler:** And he had that insight so early and understood that all he needed to do was create that space. 'Cause if you don't, and you don't create that space early, it's gonna fragment. You're gonna get, like, eight of these channels or a bunch of pockets of it. But by creating the one surface area and owning it and being synonymous with how are we gonna figure out AI for the company and the space where people are chatting about it, he was then able to just add all of this jet fuel to all of the other things he was working on.

\[00:33:41\] **Andrew Zigler:** Because you buy... You get all this bought-in implicit trust from leaders when they join ch- a channel, and it's like you're the person there or curating it or the first person to show up on the topic. So, like, if you have an open Slack in your company or Teams or whatever, and you think there's a topic that's gonna be top of \[00:34:00\] mind for your company in the next, like, half year or year, and you wanna work towards it, literally go make a channel.

\[00:34:05\] **Andrew Zigler:** Just go own that problem and start the communication there. Work on it in the open, and that's how you start to build that cross-team, uh, trust

\[00:34:13\] **Ben Lloyd Pearson:** All right. Well, we've been talking about software factories a lot today, and I think we're gonna keep talking about it. And of course, I wanna just remind all of our listeners, if you haven't seen it yet, we just hosted this wonderful software factory roundtable with some of the leading minds in, uh, the agentic SDLC. If you wanna go check it out now, just head right over to LinearB website. You can actually go get early access to the on-demand version of it, uh, if you weren't able to make it to our live event. and then of course, i- if you'd rather just wait, it'll also come out on Dev Interrupted, so make sure you're, your eyes open for that one. So, thanks to all of our listeners for checking in with us this week, sticking around all the way here till the end. reach out to us on social media. We're on

\[00:34:57\] **Andrew Zigler:** Please.

\[00:34:57\] **Ben Lloyd Pearson:** On Substack. Um, engage with us. You know, we \[00:35:00\] love to get comments and to have discussion with, with the people that are out there consuming our content.

\[00:35:04\] **Ben Lloyd Pearson:** So thanks for joining us this week, and we'll see you next time.

\[00:35:07\] **Andrew Zigler:** See you next time.

## Real conversations with top engineering leaders

Find us on

[](https://www.linkedin.com/showcase/dev-interrupted/)
[](https://devinterrupted.substack.com/)

## Your next listen

[![Cover image for Can agents keep a secret? We asked 1Password’s CTO Nancy Wang](https://assets.linearb.io/image/upload/c_limit,w_3840/f_auto/q_auto/v1/Blog_Comprehensive_DORA_Guide_2400x1256_78_8ed56543e1?_a=BAVMn6ID0)](https://linearb.io/dev-interrupted/podcast/1password-nancy-wang-agentic-security-secrets)

Dev Interrupted

[Can agents keep a secret? We asked 1Password’s CTO Nancy Wang](https://linearb.io/dev-interrupted/podcast/1password-nancy-wang-agentic-security-secrets)

1Password CTO Nancy Wang joins the show to break down the golden path for agentic security, explaining how just-in-time secrets provide autonomous AI with...

[![Cover image for The battle to replace Github, building assembly lines for software, and why no one finishes projects anymore](https://assets.linearb.io/image/upload/c_limit,w_3840/f_auto/q_auto/v1/the_battle_to_replace_github_building_assembly_lines_for_software_and_why_no_one_finishes_projects_anymore_cbbc64ca4a?_a=BAVMn6ID0)](https://linearb.io/dev-interrupted/podcast/github-outages-software-factories-agentic-assembly-lines)

Dev Interrupted

[The battle to replace Github, building assembly lines for software, and why no one finishes projects anymore](https://linearb.io/dev-interrupted/podcast/github-outages-software-factories-agentic-assembly-lines)

This week on the Friday Deploy, Ben and Andrew unpack how surging agentic code volume and persistent GitHub outages are challenging traditional code hosting....

[![Cover image for Agent, skill, or MCP? Which to use and when to use them | AWS’ Clare Liguori](https://assets.linearb.io/image/upload/c_limit,w_3840/f_auto/q_auto/v1/Blog_Comprehensive_DORA_Guide_2400x1256_76_bb95f7500e?_a=BAVMn6ID0)](https://linearb.io/dev-interrupted/podcast/aws-clare-liguori-agent-skill-mcp-architecture)

Dev Interrupted

[Agent, skill, or MCP? Which to use and when to use them | AWS’ Clare Liguori](https://linearb.io/dev-interrupted/podcast/aws-clare-liguori-agent-skill-mcp-architecture)

AWS Senior Principal Engineer Clare Liguori joins the show to untangle modern agentic architecture and help engineering teams choose between full agents,...

## Structured data

_Machine-readable metadata (JSON-LD) embedded in the page for search/AI context — not content rendered on the page itself._

```json
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "LinearB",
  "url": "https://linearb.io/",
  "logo": "https://assets.linearb.io/image/upload/v1715628027/logo-mark-lg.svg",
  "description": "LinearB is the engineering productivity platform that helps engineering leaders prove AI is improving throughput without sacrificing delivery confidence, flow efficiency, or developer experience.",
  "sameAs": [
    "https://www.linkedin.com/company/linearb"
  ],
  "award": [
    {
      "@type": "Award",
      "name": "LinearB is a Leader in the 2026 Gartner® Magic Quadrant™ for Developer Productivity Insight Platforms",
      "dateAwarded": "2026",
      "awardedBy": {
        "@type": "Organization",
        "name": "Gartner®"
      }
    },
    {
      "@type": "Award",
      "name": "Great Place to Work Certification",
      "dateAwarded": "2025-2027",
      "awardedBy": {
        "@type": "Organization",
        "name": "Great Place to Work"
      }
    },
    {
      "@type": "Award",
      "name": "America's Best Startup Employers 2025",
      "dateAwarded": "2025",
      "awardedBy": {
        "@type": "Organization",
        "name": "Forbes Magazine"
      }
    }
  ],
  "hasCertification": [
    {
      "@type": "Certification",
      "name": "SOC 1 Type 2"
    },
    {
      "@type": "Certification",
      "name": "SOC 2 Type 2"
    },
    {
      "@type": "Certification",
      "name": "GDPR Compliance certification"
    },
    {
      "@type": "Certification",
      "name": "ISO 27001"
    }
  ]
}
```

## More on linearb.io

### Top navigation

- [Book a Demo](https://linearb.io/book-a-demo)
- [AI Code Reviews — Catch security risks, bugs, and spec mismatches](https://linearb.io/platform/ai-code-reviews)
- [AI & Productivity Insights — See how AI tools affect cycle time and delivery speed](https://linearb.io/platform/ai-developer-productivity-insights)
- [Measure AI Impact — Track AI adoption and tie it to delivery outcomes](https://linearb.io/use-case/measure-ai-impact)
- [MCP Server — Chat with your data to spot patterns and boost output](https://linearb.io/platform/mcp-server)
- [Resource Allocation — Cost initiatives and shape your investment strategy](https://linearb.io/platform/resource-allocation)
- [Cost Capitalization — Capitalize engineering costs with audit-ready reports](https://linearb.io/platform/cost-capitalization)
- [Dev Team Management — Set targets and tie throughput to business outcomes](https://linearb.io/platform/goals-and-reporting)
- [DevOps Workflow Automation — Policy-based PR routing, approvals, and tests](https://linearb.io/platform/ai-workflow-governance)
- [AI Powered Support — Unify AI and human code delivery in one clear view](https://linearb.io/use-case/ai-powered-support)
- [Optimization — Surface friction with feedback and MCP insights](https://linearb.io/platform/developer-experience)
- [Reporting — Spot what's working and what needs attention](https://linearb.io/use-case/measuring-developer-experience)
- [Surveys — Turn developer feedback into actionable signals](https://linearb.io/platform/developer-surveys)
- [Platform overview](https://linearb.io/platform/overview)
- [Watch now](https://linearb.io/resources/the-great-software-factory-debate)
- [Customers](https://linearb.io/customers)
- [Pricing](https://linearb.io/pricing)
- [Why choose LinearB — Explore your data. Measure performance. Act to improve it.](https://linearb.io/why-linearb)
- [APEX framework — The operating model for AI-era engineering teams](https://linearb.io/resources/apex-framework)
- [Anti-FAQ — The questions other vendors won't answer](https://linearb.io/why-linearb/anti-faq)
- [Security — Enterprise-grade compliance and zero code access](https://linearb.io/security)
- [Build vs. buy — The hidden cost of building it yourself](https://linearb.io/resources/build-vs-buy)
- [Dev Interrupted Podcast — Conversations with engineering leaders](https://linearb.io/dev-interrupted/podcasts)
- [Reports & Guides — Deep dives on productivity and delivery](https://linearb.io/resources)
- [Webinars — Expert sessions on productivity and AI](https://linearb.io/resources?category=workshops)
- [Metrics Benchmarks — See how your engineering org stacks up](https://linearb.io/resources/software-engineering-benchmarks-report)
- [Blog — Product updates and practical insights](https://linearb.io/blog)
- [Help Center — Documentation, setup, and support](https://linearb.helpdocs.io)
- [API Docs](https://docs.linearb.io/api-overview)
- [Status](https://www.linearbstatus.com/)
- [Integrations](https://linearb.io/integrations)
- [LinearB Library](https://linearb.io/library)
- [Engineering metrics](https://linearb.io/library/engineering-metrics)
- [Platform engineering](https://linearb.io/library/platform-engineering)
- [Engineering glossary](https://linearb.io/library/engineering-glossary)
- [Developer productivity](https://linearb.io/library/developer-productivity)
- [AI in software development](https://linearb.io/library/ai-in-software-development)
- [Engineering management](https://linearb.io/library/engineering-management)
- [Developer experience](https://linearb.io/library/developer-experience)
- [DevOps](https://linearb.io/library/devops)
- [Engineering operations and the context layer](https://linearb.io/library/engineering-operations)
- [Engineering efficiency](https://linearb.io/library/engineering-efficiency)
- [Software delivery](https://linearb.io/library/software-delivery)
- [Research and data](https://linearb.io/library/engineering-benchmarks-and-research)
- [LinearB is a Leader in the 2026 Gartner® Magic Quadrant™ for Developer Productivity Insight Platforms](https://linearb.io/resources/gartner-magic-quadrant-dpi-platforms-2026)
- [Sign in](https://app.linearb.io/login)
- [Enterprise](https://linearb.io/solutions/enterprise)
- [Contact](https://linearb.io/contact-us)
- [About us](https://linearb.io/about-us)
- [Careers](https://linearb.io/careers)
- [Service agreement](https://linearb.io/services-agreement)
- [Privacy policy](https://linearb.io/privacy-policy)
- [DPA](https://linearb.io/data-processing-agreement)
- [Security FAQ](https://linearb.io/security-faq)
- [Substack](https://devinterrupted.substack.com/)

### Footer

_Additional links from the site footer, not repeated from the top navigation above._

- [GitHub](https://github.com/linear-b)
- [LinkedIn](https://www.linkedin.com/company/linearb)
- [Twitter](https://twitter.com/LinearB_Inc)