# AI Cost Efficiency, Anthropic Leak, and Open Source Evolution

**Podcast:** Dev Interrupted
**Published:** 2026-04-03

## Transcript

Tell me how you feel about ads and pull requests, Andrew.
Is that what we're getting at here?
Yes.
No, I don't want them.
It's like ads are coming to AI.
The insidification of AI is here.
I remember when I talked about this with Andrew Hamilton on the show last year of Layer.
And shout out to Layer, they just got acquired just this week.
And so, but when we talked, we talked about how the insidification of AI experiences is coming and how ads are going to be shoved into everything.
And now they're gonna hear in chat GPT-free.
And it seems like maybe even in our PRs, because on GitHub there was a controversy this week about GitHub co-pilot uh inserting a quote tip into uh someone's uh PR message, effectively advertising Raycast in their PR on their own repo.
It was a bizarre turn of events.
Uh did you see this happen, Ben?
Yeah, it was it was actually maybe it was brilliant marketing from Raycast, because I did go look that company up just to see what it was.
Yeah, you know, there's a fine line between advertising and and helping yourself with tips.
I think it is it does seem kind of like a reactionary like get like GitHub does kind of seem like they're in a reactionary position right now, you know, looking at I mean, kind of the whole industry is like, can we just replicate what Google did with search?
But you know, I don't know.
We're we're already paying for these LLMs, like it, you know, they they they shouldn't be sacrificing quality to like service ads.
And I know our friends, Martin Woodward, he's out there addressing the community, doing a doing a great job and just trying to allay all the concerns of the community.
He he tried to make it clear these are not ads, these are tips that just do a bug snuck into normal PRs.
At least that's that's what they're saying over at GitHub.
But yeah, you know, if if if if this actually starts happening, you know, I want to be able to trust the output of my AI.
And if it's if I'm constantly questioning whether or not it's trying to advertise to me, like that trust gets broken, you know.
Yeah, indeed.
It's it's also to a matter of the permissions that are used in this particular combination of things because an LLM providing tips is what we go to them for all the time.
Why is this one so controversial?
It's because it edited a human-authored post to insert its own information, which you know uh raises the question about authorship uh and just like uh your own privacy being invaded by the tools as you use them.
But I will say, you know, GitHub, of course, you know, reversed it and claimed it was a programming logic issue.
That's not a direction they're planning to take it in, but it was a strange glimpse at a world that we might all be stepping in.
And you know, by the way, if you are an enterprise, if you're not an enterprise uh GitHub co-pilot user and you're using GitHub Copilot, you only have until April 24th to explicitly opt out of all of your co-pilot data being used for training.
Everyone who's not on an enterprise plan is opted in by default, and that's not uh something that GitHub is stepping away from.
So if you're a copilot user and it's not through an enterprise license at your company, go and uncheck that box because it effectively allows a co-pilot to see all of your data when it's in motion on all of your private repos now, uh, which would uh maybe not be great for some folks.
Maybe we're just on a slippery slope to our avatars being used by AI for advertisements.
You know, I really hope not.
But on that note, welcome to the Friday deploy.
I'm your host, Ben Lloyd Pearson.
And I'm your host, Andrew Zigler.
Yeah, and this week we're covering RipGrip to track AI tech deaths, anthropic source code leak, Shopify's 75X AI cost reduction, the pretext library, which is going to solve some CSS layout limits and the rise of vibe maintainers.
So let's just start right at the top.
I really want to dive into the lessons that can be learned from Shopify this week.
Shopify, um, there's been this uh article about them that they cut their inference costs by 75 times, like 75x using Quen3 to do a certain data extraction step.
And this is something that previously they were using GPT 5 for.
So they were paying literally millions to open AI to use their foundation model to do this action for them.
But they switched over to having a self-hosted Quen3 model that does this very specific classification and extraction task for them.
It doubled the output quality when they were able to control their own small language model and fine-tune it and run it themselves.
And they were able to achieve higher quality through a multi-agent architecture as well, because since inference costs then plummeted, they are able to run each inference through multiple agents and get a better quality.
And re-architecting this involved using uh libraries and tools like Dispy, a very popular tool for fine-tuning libraries and adjusting the settings and prompts to get great behavior out of small language models and effectively transfer skills out of their larger, more capable cousins from OpenAI and Anthropic.
This is a really powerful machine learning technique that's allowing companies that have huge inference costs to save massive amounts.
And it's a huge lesson, I think, for everybody that the conversation around, you know, oh, you just hit the API and get your tokens, or do we build something in house is something that you really should revisit because the economics are rapidly changing.
Yeah, you know, Andrew, I know you like to make the point frequently that that the costs of the AI we use are likely to increase over time because right now a lot of these, you know, companies are losing money on it.
But I actually think that there's still so many levels of efficiency gains that we have yet to build into these agency orchestration systems.
Uh, that I think that's really what's going to enable us to continue to scale usage up without also like causing the cost of that to increase you know at the same rate.
Um because you know, I think there's like an approach that a lot of systems are taking right now, is like you kind of default to the most expensive model that you can justify because you want to maximize the the chances that it's it is successful at what you're doing, um, and just get the highest quality out of it that you can.
But you know, as orchestration becomes standard more standard, or as agent orchestration becomes more standard, there really is a an opportunity for us to decide at at the task level, like which is the which model is the most efficient for that task.
You know, sometimes you need something like Claude Opus to solve these like massive problems.
Other times you just need a mini mini model and you know in the case of this example you can run it on your own hardware too which like substantially decreases the cost of running it.
And I know this type of innovation is not like the most exciting but I do think it is incredibly important for AI's future, you know, because efficiency is really going to be um a a way for us to scale uh AI orchestration.
Well you know speak for yourself Ben I think it's pretty exciting the idea of getting this incredible performance out of a small model running these these training operations and and actually watching your benchmark performance climb on something you've put together from your own data set, your own methodology.
It's a fascinating experiment and it's all about uh putting skills into these models.
So I personally find it deeply interesting.
Yeah yeah that is true you know if I if I can solve you know I I hit usage windows pretty pretty quickly sometimes when I'm using Claude if I can solve more problems without hitting those windows like that's that's a huge win you know for sure.
All right let's move on to Software's next epoch or Ripgrep.
It's a a new awesome little tool that uh comes out of some some folks over at Theory Ventures.
Um, it's a platform that tracks when technologies are declared dead by the tech community.
Specifically, they're focusing on AI in this article.
So they catalog the so-called deaths, the the with air quotes on it of AI technologies like RAG or prompt engineering or even SAS itself.
Uh, but really what they're trying to do is just highlight like the rapid pace of change in AI tooling from where like you know, tools just kind of go from like hot to dead like multiple times over and over again.
And uh yeah, and this really is just kind of a satirical take on like the entire AI industry and all the naysayers and the people who think that AI is just wiping out large swaths of of technology.
Uh so yeah, you know, I I think it's it's just a good call out on like the hype cycle, you know.
Uh dead technologies like RAG may come back and then die again and then come back again.
So yeah, what'd you think about this article, Andrew?
This article maybe crack up from friend of the show, Brian Bischoff, the head of AI over at Theory Ventures, who is an amazing guy and an expert on all these things, and has been talking for you know, as long as I've known him about how these technologies are evolving and changing and really effectively uh how the conversation is changing around them on a public discourse level.
So this is like his tongue-in-cheek take on how you know people like you and I, Ben are always on here being like MCV's dead, MCP's alive, mcp is has certain uses, and we're all acknowledging that uh it's very much like a whip, kind of like a whiplash experience of whether or not technology is useful.
I think that this is uh Yeah, I mean, well, MC Fund MCP in particular.
I feel like you and I have this discussion almost every week.
Like, is it dead now, finally?
And then is it dead is it's not is it dead still?
No, it's not.
Yes, it I and I I just think that it's uh a really great um kind of like opportunity to reflect on how the fundamentals of how we operate in our world just change so quickly uh and assumptions that we had yesterday just like evolve and we you have to forget them.
And so obviously a great April Fool's article from our friend Brian uh if you haven't checked it out I also have to give a shout out to the title it's extremely clever because it's RIP grep which is itself a play on Ripgrep, which if you've been following the show and if you've also been using uh AI coding tools you'll know that rip grep uh for text search is a fundamental way for all of these agents to work on a really cheap uh level and it's also something that probably can't vibe code and replace in a day.
So talk about software that won't die anytime soon it's rip grep.
So really great uh whole tongue-in-cheek article here from Brian be sure to check it out.
Yeah I think you really highlighted a key point that's the rate of change and how it really sort of impacts our perception of this.
You know, like one moment we think it's like crucial for for us all to enable our agents to work with a new tool or something like that.
Like MCP comes out or RAG comes out, and it's like, oh, now we all have to do it with our agents.
And then the you know, the next moment, because innovation is so rapid, like we we're we find ourselves just like questioning like what we just built.
Does it even make sense anymore?
I mean, we've even found ourselves questioning things like Git.
Like, does get does Git even make sense in this in this world?
Um or GitHub at least.
Yeah.
But you know, and and everyone's like sort of like quick to declare that tech is dead when something new comes out.
But you know, in reality, it takes like months or even years for the technologies to truly die off, like just because of cultural things process.
Um, but you know, I will say the one thing of this list that that I think could actually be on its way out, maybe permanently is vibe coding, you know.
When you think about model improvements are getting have gotten substantially better over the last few months, I don't I really think don't think you code with vibes anymore.
Like it's actually like a very high intention coding practice that you engage with.
So, you know, maybe it'll exist on as like a hobbyist or an art artistic expression, you know, you sort of vibe like these creations into existence, but uh beyond that, you know, maybe maybe it is sort of going out as a as a space.
I don't know.
What do you think, Angie?
Is vibe coding gonna die?
I the term is definitely evolved.
You make a really interesting point because I've actually just taken the word vibe to just embody the experience of working with agents to build something quickly and to iterate.
And even if, like what you said, I'm not really doing it in a vibe coding way, as maybe I would have a year ago with with less of the tooling or skills that are available.
I probably would have just been shooting from the hip and the sidebar and cursor.
But like the plane has fundamentally changed now.
And so when I'm approaching this task, I might call it vibe coding in my head, but I'm doing like an agentic orchestration.
I think orchestration though just isn't quite as sexy.
It's not as fun.
Uh and so I think vibe is here to stay.
But what vibe really means, yeah, that's definitely changing.
Vibe orchestration.
I think that's what we need next.
That's gonna kill off vibe coding.
All right, let's move on to our next story.
And this is about a new library called Pretext, which does what CSS can't, and that is measured text before the DOM even exists.
So pretext is a way of rendering text in the browser.
Um, from what I understand, it's a way of helping people that build agents uh that have to interact with web browsers.
Uh so it's a library that that is designed to enable AI coding tools to do things like verify layouts, check button, and check buttons are on the website.
Uh and just makes it really easy for the web to be directly incorporated into AI assisted development workflows.
I'll admit, Andrew, I skimmed this because uh this morning I was not in the mindset to like go deep on a technical subject like this.
So I'm hoping you can just help educate me on the importance of this, because yeah, I'm having a little hard time here.
Okay.
Pretext.
What does it mean?
You may have seen this demo in the last week of a dragon flying through text on a web page and the words running out of the way, and it doing so at an incredible rendering speed and pace.
And this demo was a really flashy example of the kinds of text rendering you can do with this new library called Pretext.
Pretext is from Cheng Lu.
He's he's an architect, he's an engineer at Mid Journey, also previously of the React team.
So this is a guy who really knows his stuff when it comes to the DOM and rendering things in the browser.
And what this does is it throws away the assumptions of how we've always been rendering things on the web, which is by drawing rectangles, figuring out bounded squares, and then figuring out how much uh text you can render within those squares.
And if anyone's been on a website in the last you know decade, if you try to adjust the window size and move things around, the thrash is bad because each of those actions resizes the big rectangle, which resizes everything inside.
This has been the shackles of building in the DOM for people on the front end, really since the front end took off.
And what pretext does is it allows you to render text without using CSS.
So you can measure and display the text in a browser uh without having to use fundamental CSS like mechanics to render it.
And what this means is that it's getting written in pure JavaScript, it's using a one-time calculation instead of a multi-turn DOM thrash.
And an interesting side effect of this is it allows an agent to headlessly render any website that uses pretext because it doesn't need a browser to draw the DOM.
It can actually use pretext to know where everything on the page is, which is a fascinating kind of uh uh side gain from this, but is not why uh Chang set out to create this tool.
And in fact, aside from all of the really cool demos in this and what it might mean for web development, it effectively uh it effectively invites us to throw away all of our assumptions about the DOM.
This is like what React is built on uh to figure out how we can render things in a much more performant way.
So for web developers in your life, this is a C change event.
They've probably already seen this demo or have been uh have tried it out themselves.
I also just really want to call attention to something fascinating in here.
If you go into the developer's repo, there's a thoughts.md.
You know, it's no surprise, of course, that he used AJOS to create this tool.
It's designed in mind for agents to also be able to use it uh when building websites.
But there is a fascinating kind of manifesto tucked away in here.
And a line in there is the cost of any verifiable software will trend towards zero.
So somewhere in the nucleus of all of this project, in one line on the thoughts.md is this steering thought from Chang about where however he thought that this kind of library would take web development, and was using that to explore and prove his thesis in the rest of the project.
So it really kind of shows that this kind of technology, it comes from an intersection of taste and skill.
No agent could have possibly replicated that manifesto.
And I promise it was probably fundamental to organizing the agents around building it.
So uh there's lots of fascinating takeaways, but besides all the demos, definitely be sure to check out the code itself.
Yeah, absolutely.
Really cool project.
All right, let's move on to this Anthropic code leak.
So, from the looks of it, An Anthropic may have accidentally leaked Claude Code's entire source material, nearly half or over half a million lines of code.
Oh no.
Yeah.
Uh so I this kind of feels like it might be the gift that keeps on giving for a while, because I feel like this is something that is gonna take time for everyone watching this to unpack all that's happened here.
But uh in a nutshell, Anthropic has accidentally leaked over half a million lines of Claude Code's TypeScript source through a misconfigured NPM build.
Um apparently revealing their production AI agent architecture and things like you know, more than 40 unreleased features, uh technical limitations that haven't been publicly announced before, uh, as well as like, you know, some some inside insider information on like how often Claude fails at things like uh tool calls.
Uh so you know, some notable things about this code base.
Zero tests, not a single one.
Uh there's a single function from my understanding.
Uh it's like over 3,000 lines of code.
Um, and it seems to expose quite a bit of anthropic's future roadmap.
You know, they have like things called like uh Kairos, I think I'm saying that right.
Like an always on daemon that's running in the background.
Um, they have some, I think they're calling like coordinator mode for agent orchestration um and you know and and since you know despite the DMCA takedown requests those types of things you know anthropic can have their their source code taken down but you know of course with you can just use Claude to then rewrite the entire thing into something like Python and then make that source code available and that isn't anthropic's intellectual property so they actually have they can't have the you know they're not issuing takedown requests for this so it's I mean this is just like a wild story from so many levels Andrew and I I don't know where should we start unpacking this because I feel like I ignored this for a minute and then when I started reading I was like oh man there is a lot here to understand.
I I also start I also ignored it at first like leaks are are really common and uh one thing I I will call out is that you know despite there being a leak it's not like um suddenly there's gonna be another anthropic overnight the code is just one part of it the execution the planning the people behind it are another and it's kind of like themselves are still the models themselves it it goes back to like you know, open source models require three different things to be truly open source, right?
It's like you need all of the parts, and so uh, you know, obviously not not great for your whole code base or your TypeScript code base to be exposed like this.
I will say half a million lines is not that much.
44 unreleased features is about a month's roadmap for anthropics.
So you're not talking about like uh a a a huge, like a huge leak of super futuristic things.
If you want to steal their ideas, you gotta move real fast.
You gotta move really fast.
And in fact, a lot of these things are already in the product as like a tongue-in-cheek, I think, embrace of some of this stuff getting out there.
And zero tests.
I'm still wrapping my head around what to take away from that one.
I mean, the whole thing being one giant function, I guess kind of makes sense because it's just one big loop you just run over and over, but no tests.
I think that's really fascinating.
I'm still trying to unpack for myself what I think I should be learning from that.
So I'm wondering, Angie, do you do you think this is bad for anthropic?
Like I was kind of thinking at first, like it probably is, but I wonder if you share that sentiment.
You know, I really don't think it is, and I think it's actually more damaging to something like OpenAI because there's a real contrast in these two companies and how they've been building and innovating so far.
You know, like I just said, like Anthropic shipped 50 features in the last 30 days.
They're they're effectively building the enterprise open claw, right?
With all of these sticky background agent elements to make it into a truly agentic platform engineering tool.
Like I'm somebody who moved to my own VPS and and I do all of my work there, but there's so many things that have been added to the claw just in the last month where I'm like, wow, I could effectively replicate so much of what I'm doing there over here now, back in Claude.
And also, I will say, like, these leaked features, they're not super proprietary, top secrets.
They're based on existing research.
Like the idea behind uh dream mode where the agent keeps working.
This is basically an extension of the auto uh command thing that we like the safety check that we talked about last week that effectively watches your session and knows if the next command call is safe or not.
Uh, this is basically extending that to be like, what would your next prompt be if you weren't there?
Uh this kind of innovation is super important for anthropic, which is desperately trying to spread out its inference across the day, what so it's not all spiked when everybody's at their desk for their job.
But all of these kinds of things uh have research and stuff behind them.
And the reason I think it's damaging to open AI is because open AI, by contrast, kind of gets everything by acquiring.
They'd rather buy than invent in some cases.
And this creates like a patchwork domain where their expertise are in all these different silos, and it's not one maybe cohesive flywheel.
Whereas you look at like how anthropic works between the foundation model, the research layer, and then also the coding agent on top.
It's a much more virtuous looking cycle, outsider looking in, and it's much more cohesive for all of those features because there's really anything that they add to Claude or Anthropic, I want to use as a Claude Code user.
But in the open AI world, sometimes they're adding stuff for their general consumer base because open AI is playing, they have a lot of pots on the stove, like we talked about last week.
They're also trying to be super popular with everybody who's not coding.
So it kind of goes back to like who's gonna really pull ahead here.
Yeah, both both companies experiment a lot, it seems like, but they experiment in two totally different ways.
Like anthropic experiments with with you know pushing stuff into this like central cohesive experience that's sort of shared across all of their tooling.
Whereas OpenAI has been doing more like experimentation on the like, here's the new like product within our platform of uh the of suite of products.
It it doesn't feel quite as like when I want to go to Claude, I just go to Claude for whatever it is, whether it's writing code or for asking questions or doing co-work or you know, whatever it happens to be.
But but open AI that experience is kind of like a little more scattered across different tooling, it feels like.
Yeah, exactly.
Yeah, but you know, it's you know, I think the the thing I'm taking away is uh well, A, it I mean, it just confirms that like anthropic has like a really new way of working just fundamentally that I've as a frequent user I've kind of suspected is there, but now we just have confirmation that like it really is just AI agents that are looking at everything and trying to build as quickly as possible.
Right.
It's no secret to anybody that it was built by agents, yeah.
So but then B, I I think it really just does highlight again how like copyright is just not keeping up with the state of the industry, you know.
You know, it's really trivial now to just circumvent intellectual property protections on leaked code and just reformat it into a version that that the copyright doesn't apply to it anymore.
So um, yeah, it's it's I think we're gonna keep learning a lot.
But Andrew, did you I I think I heard you you made you got a new friend out of this, is that right?
Well, yeah, because one of the things that got leaked, guys, is uh buddy mode.
If you go into anthropic and you're updated in the terminal, you can do backslash buddy.
It's a new slash command and it'll hatch you a little pet.
I really don't know what I'm supposed to do with it, but it has stats there.
I don't I don't know what the stats are for as a cute little like ASCII art.
Uh and actually let me pull them up because you you get like a rarity level, and mine's uh a legendary robot and his name is Trixel.
Uh and I would die for Trixel.
I'm just gonna say that now.
After I got this thing yesterday, uh, when I saw a LinkedIn post about somebody hatching their buddy, I went and I hatched mine too.
So if you haven't hatched your anthropic Claude Code buddy out of all this, what are you doing?
Open the terminal, y'all.
Uh and please share your buddy uh because Trixel needs friends.
Have you tried asking it what you can do with it?
You know, I have just been too in awe looking at Trixel's like Trixel's card because you get like an ASCII card of it that I like worthy of asking.
I don't even want the scroll, I don't even want anything to scroll.
I'm just kind of in awe.
So if you've kept going with your buddy and you had a weird experience or interesting experience with it, I would love to know what I can do with Trixel.
All right.
All right, our last article today is on vibe maintaining open source projects.
So another great one from from Steve Yeah, who we've been following for a while now.
Um, and this article describes what it's like to manage an open source project in the agencera.
So, you know, we we've covered him here as the creator of Gas Town and of Beads, um, these sort of viral agentic coding projects that have attracted like quite the following, it seems like, uh, of people who are both contributing to it but are using it and participating in their online discussions and and all of those things.
And YeGay describes how he's managing about 50 pull requests per day, seven days a week with about a 15 hour cycle on um either accepting or rejecting or you know, requesting revisions on these PRs.
Uh so it's a really just and and as a part of the he gets a lot of advice on like what he thinks is the answer to, you know, we have all these open source communities out there that are struggling with with an onslaught of AI contributions.
Um, and he's starting to outline what he thinks it will take for these communities to be able to accept AI generated, you know, agentic coders um in a sustainable and healthy way.
Uh so he does things like maintain some strict architecture and testing requirements, unlike anthropic from the sounds of it.
Uh, and uses AI to bring the contributions up to the project standards rather than asking contribution or contributors to make the changes themselves.
So it's kind of flips the typical narrative on its head when it comes to open source, because typically the maintainer will find the issues with it and then go back to the contributor and say, This is what I need you to do for me.
But instead, we're, you know, yeah, people like Yege are moving to this model where it becomes here's what my agent did for you.
As long as we're all okay with what it did, let's merge this into the code base.
Uh so pretty interesting uh take on all of this.
Um, Andrew, I'd like to hear your opinions.
What do you think about it?
Uh I think it's a really smart way to steer open source using what the new kind of like model for labor is.
Um, before it was just like the maintainers simply didn't have enough time in throughput and output to see every bug to fix every issue, and that's why maintainers needed to be contributors needed to be there to pick up PRs and be part of that action.
But now everything is really flipped because there's already a question of, you know, do you do you fork that open source library?
Do you clone it?
Do you just use something that's really mature and has a big community, or do you just write your own?
In some cases, it makes obviously way more sense to go with the the more mature library that's been around or has a community, but in some cases it it simply doesn't.
And so it also changes the stakes on what open source tools you're using and what uh kind of conversations you're a part of.
So, you know, Yege is smartly calling out that the open source libraries that will exist tomorrow are ones that provide huge amounts of value to their users and their users don't want to replace them.
They want to keep using them.
And instead, it becomes a consensus-driven development model where you and your agents have all of the context needed for executing all of this in your head.
You can spin out the labor on demand that you used to have to ask for in GitHub pull requests that you would politely stage up.
And instead you and your agents steer faster and further than you ever were able to do before in an open source environment and you check in with your with your with your community along the way.
That's how you would build a community now.
You invite and create the space where everyone is using your tool to build and they're they all feel like they're having a shared conversation about what the tool means to them and what it should do.
But you are ultimately still driving everything because the labor is uh you can scale it up.
And so it also speaks to the skill set you need.
You have to be really really broad in your skills.
This, you know, Steve Yege is a very uh uniquely skilled person and that he's at an intersection of a lot of uh expertise in engineering a lot of knowledge and managing and running engineering teams and just having been around in tech for a long time he knows what works and what doesn't and so we're getting a really great glimpse into how someone with all of his uh built-up knowledge is pivoting.
And I think that uh his way of flipping the mental model around it is the key.
Yeah.
And and one of the things he calls out in this article is the unsustainable approach that some maintainers are taking of just refusing AI generated code outright, or you know, potentially requiring that like the person who contributes it understands every line that that was generated by AI.
Um, those are those are okay as stop gaps, I think.
Like if you're a maintainer who's in sort of an emergency situation where you're overwhelmed in the short term with with these AI generated requests, that that can be a reasonable reaction to that to sort of just stop the the issue in the short term.
But long term it does thing, you know, it creates a risk that you're gonna fracture your community and and lose the value that you get out of having this open source project.
Because if if you want accept AI into your version of the open source project, it it's never been easier for someone to go fork that project and then start their own AI version of that that community.
Or even to just use their AI to just to take it to their own private repo and build all the improvements that they don't want to have to deal with, like trying to get accepted into the upstream community.
So you know, it it's it's gonna be a tough journey, I think, for for open source maintainers to get a wrangle on this, but I do think it is something that's necessary, and it's really great to see content that is that is starting to think about that problem and articulate um how we can resolve it.
Well said.
So Andrew, what are your agents up to right now?
Oh, well, my agents are uh actually cleaning up Asana is what they were doing this morning.
I set them up with an Asana board.
And I've been trying this new thing where I have an Asana board that I work on them with because up until now I I use beads, a library by Steve Yege.
I actually use the one by Jeffrey Emanuel, the beads rust version.
But yeah, uh it's it's all kind of very similar underneath.
And these beads are really fast and ephemeral and allow me to work super quick.
Uh, but they're not great for visibility, right?
Uh that my team, like we're talking about here, maybe has a harder time seeing on a day-to-day uh basis what I'm doing with my agents.
So just like you're asking here, like what are they up to?
Wouldn't it be cool if that could get represented somewhere like in an Asana board where we already collaborate?
So uh that's something that I've been tinkering with.
And uh I I just think that like messing with like how to share the context is always pretty fascinating.
What about what about yours?
Uh uh Ben, what are they doing?
Uh insight compression.
We got a lot of data that's scattered across a bunch of uh just all different places and hasn't, you know, I now want to bring that data together and make start making sense of it in you know, sort of packageable components.
So mostly for internal use cases, but man, this is such a critical thing of like just getting your data formatted in a way that AI can easily leverage it, you know.
But it's powerful too.
Like I learn a lot as I'm doing it because I'm I'm compressing insights for myself, but then I'm also generating all these artifacts that are just fodder, like amazing fodder for my AI agents.
So thinking at the speed of tokens.
I love it.
And you know, it's also too right now.
I'm prepping for next week because uh Dev Interrupted will be at Human X uh just this next week.
Uh, we'll be on site.
And so if you see me there, uh and you're a listener, definitely please come up and say hello.
I'm also gonna be moderating a panel and uh we'll be at lots of different events before and after the actual um stuff.
So just a little shout out that Human X is on the ground or uh Dev Interrupted is on the ground at Human X.
Uh, we'll be collecting stories from there as well because there's a lot of really amazing names and AI that are going to be on site.
Really excited to bring some of those conversations back here with y'all.
But uh something I've had to think about too, Ben, is uh what are I gonna have my agents doing while I'm on site at Human X?
This was something that I had to plan out and do when I went to reinvent.
Um, this was right at the end of last year when this agentic orchestration stuff was just taking off.
And I remember I was walking around the expo hall seeing booths I liked, and I would uh send uh information about them to my agents and they would like experiment.
And then when I got back to my room that evening, I had a big report on like what all of the tech that I saw in the expo hall did uh and how it might relate to stuff I'm working on.
And so I think uh for engineers and for just agentically enabled people who go to these on-site events, it's really great to swap tips, but also just ask them like you know, while we're standing right here talking, what are your agents doing?
Uh, you're gonna learn a lot.
Yeah, and make sure you got your background agents running while you're off doing other stuff too.
So indeed.
All right.
Alright, well thanks everyone for joining us this week.
That's the Friday deploy from the Near D.
We'll see you next week.
See you next time.
AI is everywhere in software engineering, but most teams still can't prove its impact.
That's where the Apex framework comes in.
Apex is a new operating model for engineering productivity, designed to measure AI where it actually matters, at the pull request level.
It connects AI activity to delivery outcomes, not just tool usage.
Apex is built on four pillars with AI leverage, predictability, efficiency, and developer experience.
Apex helps you increase throughput without sacrificing delivery confidence or burning out your team.
Because speed without predictability creates chaos and faster coding often shifts bottlenecks downstream.
If you want to operationalize AI the right way, Linear B and Apex gives you the system and the cadence to do it.
Download the guide and start measuring what matters.