The GitHub Problem  ## Document Notes ## Highlights - **Glasses Changed How They See The World** - Jerod and Jared shared personal vision-correction stories about getting glasses and how it changed perception. - The hosts used these anecdotes to segue into design and tooling clarity metaphors for tech improvements. Transcript: Adam Stacoviak You know never asked don't ask don't tell is my policy okay i guess so that's my laser policy as well that's a sweet story though um yeah that is kind of beautiful really uh but i don't think Mat Ryer You are anywhere near Adam's sad day. No. Yeah, ([Time 0:10:20](https://share.snipd.com/snip/5f954de5-9143-4906-825a-fc0c6c879782)) - **Rob Pike's AI Christmas Backlash** - Rob Pike received a heartfelt AI-generated Christmas email and publicly erupted in anger at being thanked by a machine. - The hosts discussed how the AI Village campaign meant well but misfired by mailing notable figures like Rob Pike at sensitive times. Transcript: Adam Stacoviak Segway. Rob Pike. Rob Pike had a very sad day or a mad day. But either way, it's an emotion, and I needed a reason to switch the conversation over to Rob Pike. Did you guys hear this? Did you guys hear what happened over our break with good old Rob Pike? Little Rob Pike. No, I said good old Rob Pike. Oh, I thought you said little Rob. I'm not going to diss the guy. No, I'm good old. He's good old. Matt, tell everybody who Rob Pike is for those who don't know. Mat Ryer Yeah, so Rob Pike has a great, in software and in computer science. He's done things like UTF-8. So I think we can thank him for emojis. That's a hit. Yeah. Yeah. He also was a co-founder of the Go Language. That's how I know him. Same. And his sensibilities and stuff. Yeah. Because I did that, that go podcast, go lime. Was it that podcast we used to do? Go lime. Jerod Santo It's somewhere. I think it was go somewhere. Mat Ryer Go away. Yeah. Um, and yeah, so, so he's brilliant. He's the plan nine and stuff. He's done loads of like important tech stuff, which is we've, the rest of the world has then built on top of as well so right it's really immeasurable impact really when you get to that Level yeah um but he's he had a bad day he had a really bad day when an ai emailed him a kind message you better get out of here yeah okay so good good so there's rob pike so adam uh similar explainer Adam Stacoviak Now, but give us AI Village. Tell us what AI Village is. Oh my gosh. Jerod Santo I didn't investigate this deeply. So give me a chance to paraphrase. From what I understand, it is essentially a village of various LLMs that have been unleashed on a desktop environment that are just being autonomous or some version of autonomous, Just doing things, acts of kindness, things like that. And they emailed the wrong person. Obviously they learned how to email these things, you know, and then now they're like, Hey, we got to prompt you back and say like, Hey, don't email people. So these things called Opus GPT three, no, GPT-3, GPT-3, GPT-5 Gemini 3 Pro, and DeepSeek v3.2, four different AI LLMs that are just thanking people, basically. And they're just autonomous, doing their thing and talking to each other. And it's an experiment, essentially. AI Village, theaidigest.org/village. Mat Ryer Yeah, their goal is to raise as much money as they can for charity. So it definitely comes from a good place. It's like these bots, because I suppose emailing, cold emailing people, does that raise money sometimes for charity? If you get an email, do you ever click through and say, yeah, I'll put my credit card details. What's the big deal? Right. Yeah. I probably don't do that, but that's because i don't want to give to charity not because i've got any data concerns about my security um no but but yeah then they said right just do random Acts of kindness and then yeah it just sent a really sort of heartfelt sounding email but of course it comes from an AI. So you know it doesn't have that. Adam Stacoviak It's not meaningful, right? Right. You said a heartfelt email, but there's no heart because there's no human, right? It's just pseudo heart. So Rob Pike receives an email. Jerod Santo Yeah. Adam Stacoviak From Claude Opus 4.5 model. The subject was awesome. The subject from AI public. Thank you for Go, Plan 9, UTTF-8, and decades of Unix innovation. Now, part of the story here that we haven't talked about yet is he received this at 5.43 a.m. So maybe he's not a morning person and he was up checking his email for some reason. And so he was mad already because he's up at 5.43, maybe. But I'm not going to read the whole email. I will say it starts with, dear Dr. Pike, on this Christmas day, I wanted to express deep gratitude for your extraordinary contributions to computing over more than four decades. And then it goes on. He did not like this. Jerod Santo You have to appreciate the fortitude, I would say, for this AI village to try to accomplish its mission. Like what a, maybe not the best day. You can train something on the world's knowledge, and then you can unleash it in a way that it can just repl through and through a reward system and accomplish to some degree or attempt To accomplish a goal. I mean, I'm not that far removed from the absolute accomplishment that this is. Now, is it ill-placed? I mean, it was smart enough or at least some version of smart to try and email on christmas day that's a day your heart is kind of open to loving the world you know loving other people wishing Mat Ryer People well you're at least merry based upon it being merry christmas right and the thing is in a way depending on how the people are, in a way it is a compliment that does have meaning. Because it essentially means, if it's just plucked it out of his brain, it essentially means that Rob Pike has had enough of an impact, as far as it's concerned, that it's worth sending An email to him. So there is at least that positive side, which you could say, yeah, you know, you could say that is actually a compliment. Jerod Santo That is not how Rob took it, okay? As you may know. Can we read this out, Jared? Do we want to have bleeps this early in the year? Let's pre-bleep it so we don't have to actually bleep it. What does that mean, pre-bleep it? I'll just read it that way. Okay, you pre-bleep it. Mat Ryer You're doing that at the time, though, not before, are you, Jared? You're doing it as you're speaking. I think that's current bleeping. Adam Stacoviak Normally we put our bleeps in in post, but I'm going to put it in in pre. Okay. Just do it live. Yes. Just in time, bleeps. Go. Bleep, you people. Raping the planet, spending trillions on toxic, unrecyclable equipment while blowing up society, yet taking the time to have your vile machines thank me for striving for simpler Software just bleep you bleep you all i can't remember the last time i was this angry yeah and he published that response for the all of us to enjoy on the internet let's look at the timestamp. Jerod Santo Okay, so 5.43 a.m. Was the timestamp according to this email screenshot. According to the published post on Blue Sky, 5.25 p.m. Same day. So he thought about this for 12 hours, potentially. Or maybe he didn't read it, guess, when it came in. I assumed he was reading it immediately. Well, I don't know about that because if he took the screenshot, which is in his post, which is at 5.43 a.m. That's the time it had to be, right? Yeah, ([Time 0:10:48](https://share.snipd.com/snip/af05b872-a6eb-4708-8af4-f4692c958978)) - **GitHub's Monoculture Risks** - GitHub's dominance created a monoculture that makes alternatives invisible and raises systemic risks when policy or outages hit. - Actions, runners, and integrations increase lock-in beyond simple Git hosting, magnifying user impact when pricing or outages change. Transcript: Adam Stacoviak Use it for good, not for bad. I'm not sure they would hear me. What do you think, Adam? Jerod Santo I just want them to stop doing what they're doing, okay? Adam Stacoviak Just stop what? What are they doing? Jerod Santo You know, just think about the platform you got, okay? Don't rug pull, not cool us. You know, don't change actions. It's like the best thing you've done in so long you know don't don't mess with it just keep it and i guess they did right they backtracked that they walked back this tell them the full story Adam Stacoviak Jared what's the full story on that one oh gosh i don't know the full story but the tldr is that microsoft announced github pricing changes around actions and the change in particular That ruffled the feathers of the hacker community was the the addition of charging for your self-hosted runners so runners that github does not themselves have to host and yet don't You dare charge for those new fees for using self-hosted runners. The people revolted. Many text areas were filled with bleeps and submitted. And they did walk it back. They walked it back. I don't know what's happening now. I think they just decided not to do it until we forget for a while. Where reddit did that one time they just like now it's a change and then everyone's like this will be terrible and then they're like well we're not going to do it right now and then they Did it like six months later and nobody noticed so maybe that's what's going to happen this time you know but i mean they should have at least done it as a pr and let everyone comment on it Mat Ryer Got merged let everyone review yeah yeah pick like jury duty that have select people that are top contributors to open source. Jerod Santo Well, they have a version of this. So they have a community org. So github.com/orgs slash community. And they have discussions. They didn't, to my knowledge, begin with a discussion, although they do say in one of the posts to let's talk about GitHub actions. And near the bottom of it, there's a headline that says, help us shape the 2026, that's the year we're in, roadmap for GitHub actions. And so they are at least in this moment, I'm not sure this is before the debacle or post the debacle, but there is some sort of request for should we sleuth it shaping it let's sleuth it what's Adam Stacoviak The publish date on that thread let's see here looks like december 11th 2025 that's pre-debacle right yes the debacle began on december 15th 2025 when they announced pricing changes For GitHub Actions. So that's four days prior. Not much time to discuss, but enough. No. Jerod Santo I think, who was it that we had on the pod that was CTO when they released Actions? He was on the pod a while back. He since stepped away. Ryan Daigle? No. Mat Ryer Join us for another exciting episode Adam Locke and Jared Change together are the ChangeLog Detectives. Adam Stacoviak Where they ask each other questions and they don't know the answers to it. Jerod Santo Leading GitHub to a $7.5 billion acquisition, Jason Warner CTO at GitHub. Jason Warner. Thank you for our search results there. To our good friends over at TypeSense. TypeSense is one of our partners to give us awesome search. Why bringing up more successful guests that you've had? Why bringing that up in front of me? Just to keep your ego in check. Contextual, man. Contextual. It's the word of the year. Word of the decade, maybe even. Context. So contextually, Jason Warner, he helped the world have GitHub actions. I think it was actually one of his brainchilds. It was this whole entire CI, CD pipeline flow that they built out was even part of the reason to acquire GitHub, you know, Microsoft's acquisition process. So that podcast covers all that. But, you know, I believe like GitHub Actions is, it's become so much so that whenever you're in your LLM, so not you building the software, your agent, your agent led, and you're building Something out and you have to do observability or you have to deploy it or whatever it might be to get it into production. The first thing it says is let's set up your GitHub actions workflow. And so it's become the default by and large for everybody. And I'm cool with charging for products, by the way. Totally cool with it. But it seemed like this was a tax on those who want to do runners externally from GitHub. And it seemed like it was like, who wants to make an announcement a week from Christmas? Like, just don't do that. ([Time 0:22:17](https://share.snipd.com/snip/74ec73b3-53d8-4379-9c95-6262f93e227f)) - **Announce Major Changes When People Are Paying Attention** - Avoid surprise product changes at low-attention times; announce big pricing or policy shifts when users can respond and discuss. - Use open community discussion and staged rollout to prevent backlash and allow contributor feedback. Transcript: Jerod Santo Don't be sleuthy. Mat Ryer Yeah. So there is that post, Jared, you sent me a post, and someone's complaining that GitHub has this monopoly. Right. Because what happens when GitHub goes down, you know, it does affect everything. It really does. The thing is, it kind of was the best choice, I think, for a long time, and it was just the easiest to use so the user experience was just kind of there and then obviously it got into loads Of integrations to it so it became it did it sort of earned its place but then it still is kind of does kind of have a monopoly a lot of Go packages and things are on GitHub, and you wouldn't Be able to pull them down, of course, if GitHub's down. Adam Stacoviak We have seen a trickle of people starting to move other places, the most noteworthy of which I think is the Zig programming language, which moved to Codeberg recently. Codeberg. Codeberg, which you bring down an iceberg. Yeah. Code. Mat Ryer Right. And it's going to bring down GitHub, the Titanic. Is that the exact metaphor? Is it? Yeah. Adam Stacoviak Nice. That's good. Yeah. And you only see that's at the tip and everything else is under the surface, you know? Right. Right. Yeah. Which you should know about, Max, is over there in the, well, they might be in the EU, which you're no longer in. Mat Ryer It was on the way, wasn't it? They left the UK and they were going to the US. They were, they were coming here. Jerod Santo A lot of controversy around that trip. Mat Ryer Not the Titanic, Codeberg, the platform. Adam Stacoviak It's a European Union platform. It is true. That's true, too. Yeah, so I guess it worked in both ways. That's not quite a double entendre. Jerod Santo It's kind of like a... That's like a triple. Yeah. Mat Ryer Especially, it's four if they mean the lettuce as well. Triple stamp and double stamp. You know there's a type of lettuce called iceberg lettuce? Do you have that in the U.S.? Oh, of course. I feel bad for the listener i'm sorry listener let's go to the iceberg lettuce go ahead i'm sorry too no i'm just saying this could be that it could be that's where they incentivize their Name like their name came from they just love a good salad because they want to be associated with lettuce they love the smell of salad if you love the smell of salad you're going to call Jerod Santo Your project something salad related maybe that's what i'm thinking i would but the main thesis i think from uh let me double check the name so i can be accurate here lionel dry cot now That's my americanized he's texas slang if that's even a thing right that's another dan tan is it if No it's not Dan Tan. Dan Tan. You gotta go Dan Tan. Who's Dan Tan? Who is he? Come on now bring it back. You'll never know. You'll never. We will never know but Lionel wrote this post and the hypothesis or at least the thesis was this was that GitHub's near total dominance over open source hosting has become a dangerous Monoculture that makes alternatives invisible, not just less popular. Kind of an interesting phrase there, invisible. Codeberg, okay, they didn't see it coming. Matt didn't even know it existed. Mat Ryer That's true, yeah. I didn't go on ships. I didn't go on many ships. Yeah. I haven't been on a ship for ages, so. Jerod Santo And you think he's a teacher or a teacher, right? He's a teacher? He's teaching students. Students couldn't, while being told to do things in open source, they only use GitHub, or a large majority use GitHub, while also being taught that there's not just GitHub. Adam Stacoviak 99%, not even a large majority, like pretty much 100%. Pretty much 100%. So I agree with you, Matt. I think it earned it. I think it earned its monopoly. It dominated for many years. I think it stagnated as well. I think both those things are true in the hands of Microsoft. I think that they don't care about it. They care about pushing Copilot into every orifice of their corporate body and into our wallets and it's gotten worse and worse and worse and worse and it will continue to do so as people Slowly move away from it until they all move away real quickly one day do the competitive products start out using github Do you think? Mat Ryer How so? When you start, what are you going to do? Make a GitHub repo. Go from there. You've got to start somewhere, aren't you? Adam Stacoviak I mean, even Go was written in C until they could self-host. Yeah. I just made that up. Is that true? I think it's true. Mat Ryer That's true. Yeah, that's true. Go was all written in C until Go was good enough that it could then be written in Go. Adam Stacoviak Exactly. Which is amazing. It is cool. I've always loved self-hosted languages, but yeah, I'm sure they have to. But the nice thing about DVCS, for those of us who like acronyms, be distributed. Is it decentralized or distributed? Version control system. I think it's decentralized. It's both. Is, I think it stands for distributed though, is that you could just have multiple origins or not origins, multiple remotes, you know, got your locals. Like there is no necessary, like, yes, your main remote is your origin is GitHub right now, but it's so easy to just change that to something else. That switching off, if you're just using Git and those flows, it's actually super straightforward. However, if you're using actions and sponsors and PRs, whatever, whatever else projects, if anybody uses that still, then it becomes much harder. Jerod Santo Yeah. It's the gravity. That's what they call it. GitHub universe. It's a, there's a gravity to it. Honestly. I mean, there really is. And so what do you do? Do you fight the system? You know, do you, if you're launching something and it's in open source, sure, you can have multiple remotes. Of course, that is totally possible. And you could, it's your prerogative. ([Time 0:27:24](https://share.snipd.com/snip/1645eb68-5a7c-4217-a0d9-a860b1eea785)) - **Tooling Gravity Trumps Protocol Freedom** - Git's distributed nature makes migration feasible, but the ecosystem and tooling gravity keep projects on GitHub. - Switching away requires tooling and user migration, not just hosting changes; ecosystem network effects are the real lock-in. Transcript: Jerod Santo Yeah. It's the gravity. That's what they call it. GitHub universe. It's a, there's a gravity to it. Honestly. I mean, there really is. And so what do you do? Do you fight the system? You know, do you, if you're launching something and it's in open source, sure, you can have multiple remotes. Of course, that is totally possible. And you could, it's your prerogative. But if the users aren't there, what's the point? Adam Stacoviak Well, it may take a tooling change to actually be significant. And I think Git is entrenched at this point because of agentic coding, especially it's entrenched. However, the further you get away from that tool, the easier it is to have your agents just go use something else you don't care anyways. And so maybe that's not always the case. However, the reason whyub became what it was was because it put all of these collaborative features around a new tool ish that was already getting popular and so they kind of popularized Each other and so maybe as jj is now jujitsu the cool new tool of the bleeding edge folks. I haven't tried it yet. I don't like to bleed as much as others, but people are loving it. Maybe a platform, not a centered around Git, a centered around JJ, which could also support all the Git things, which is currently the way JJ rolls, right? Like it's like a super set, I think, of Git's abilities. ([Time 0:32:56](https://share.snipd.com/snip/7315f964-3270-4303-8256-0e3fab9efe90)) - **Add A Git Push Load-Balancer For Redundancy** - Consider a multi-host 'pre-GitHub load balancer' or proxy to replicate pushes to alternative remotes for redundancy. - Design backup webhooks or mirrors to reduce single-point-of-failure deployment risk when GitHub is unavailable. Transcript: Jerod Santo What if we had a pre GitHub load balancer where you committed to a thing that was meant to be a load balancer? Your primary host is GitHub, but you have it. You still have downtime, of course, but you're not on GitHub's downtime. And it's not monopolized by GitHub. So what if you had a load balancer that you pushed to, but ultimately the transaction landed at GitHub to run your actions and your runners and all those things. But what if it could also webhook somewhere else and say, hey, runners over here can still do it where that way if GitHub is down, you have sort of one layer before that, kind of like the Load balancer effect to your Git pushes i don't see why you couldn't do that get like a proxy that pushes to multiple places it's all in sync it'd be a fight against it at least it'd give Adam Stacoviak Us a chance if downtime is your major concern yeah well my major concern is in downtime it's it's crappy time i don't know you call it it It's just like their bloated JavaScript UI, the Fact that like copilots shudded my face everywhere. Stuff like that is what bothers me. Like actually using the website is slow now. It used to be super fast. It's like way too much React code, I think. Those are the things that bother me more so than GitHub being down. But obviously that is a problem and we can't deploy when GitHub's down either because we don't have this proxy that you're talking about. But I'm certain that it's a good idea and that they probably exist out there. Jerod Santo Like a green blue almost. In a way. Let's say a green blue to your Git repositories. ([Time 0:44:07](https://share.snipd.com/snip/eccfa34c-5021-4842-8789-79a27485abc9)) - **2026: Agents, Vectors, And Observability** - Tomas Tunguz's 12 predictions highlight vector databases, agent-first designs, and agent observability as near-term structural shifts. - These items frame 2026 as a year where agent tooling and infrastructure become primary engineering concerns. Transcript: Adam Stacoviak Nine. So Tom has written 12 predictions for 2026. We thought it'd be fun since it is the new year and Matt has all these resolutions. I'm still rocking the same resolution I had last year, which is 1512 by 982. Mat Ryer Yes. Adam Stacoviak It's classic. Classic joke. You could have done that low-res eyes before you had your eyes fixed when you first saw a leaf for the first time in your life. And you were like, oh my gosh, there's more pixels here. That's like when we got the retina displays for the first time. We were like, wow. Mat Ryer Yeah, it's true. That's what life looks like. The text was really sharp. Adam Stacoviak It really did. And we thought it would be fun to go through some of these. We don't have to hit all 12 but uh a few that we think are either right or wrong or we have comments you know if you have comments uh sound off as i read them how should we do this should i go i'll Jerod Santo Read all 12 and then we'll stop i'll pause if you guys have anything to say matt say nothing read all 12 jared if not okay you see i preempted that here matt say nothing well he just said to Mat Ryer Talk. Jerod Santo That's why I was going to talk. Well, it's my best. No more talking for you for just a minute here. Go ahead and see them all. You prefer when you sing. Adam Stacoviak Your songs are great. Yes. But commentary, you know. Jerod Santo For the listeners' sake. Adam Stacoviak Like, if we could have a swap, if we could have a mid-show swap, you know, like Matt, if you had a tag team, maybe Tom Wilkie could be your tag team. Jerod Santo And like Tom could talk and then you could sing. That's a good idea. Let's swap out Tom next time. Mat Ryer No, the other person could be worse. Jerod Santo Just right in the middle. No announcements. Mat Ryer Back to the devil you know. All right. Adam Stacoviak So speaking of the devil, what we know and what we think we know, Tom thinks that in 2026, businesses will pay more for AI agents than people for the first time. Pausing for comments. Jerod Santo Oh, I thought you were going to read all 12. Adam Stacoviak Oh, I thought I was going to read each one and see if you guys comment. I'll read all 12. Here I go. Yeah, go fast. Go fast. Number two, 2026 becomes a record year for liquidity. That's very exciting. Number three, vector databases resurge as essential infrastructure in the AI stack. Not to be confused with the Atom stack. Number four, AI models execute tasks autonomously for longer than a workday. Dang. They'll be working hard for their living. Number five, AI budgets receive scrutiny for the first time. Number six, Google distances itself from competitors via breadth. Breadth in AI. I'm not sure what that means. Number seven, agent observability becomes the most competitive layer of the inference stack. Okay. Number eight, 30% of international payments issued via stablecoin by December. I don't know why you had to put December in there. We know the year. Okay. Number nine, agent data access patterns stress and break existing databases. I can definitely see that one coming true. Number 10, the data center build out reaches 3.5% of US GDP in 2026. And number 11, the web flips to agent first design. Say it ain't so, say it ain't so. Number 12 and the final prediction of Tomas Tungus is Cloudflare becomes the gatekeeper for agentic payments. There you have 12 predictions for 2026. ([Time 0:51:30](https://share.snipd.com/snip/98dda346-0e49-42bb-a876-4708508f1313)) - **What Agent-First Web Design Implies** - 'Agent-first' design means building interfaces with agents as primary clients, not humans, which reorients APIs, error messages, and flows. - The hosts expect agent-aware design to grow but debate whether full agent-first web design will arrive in 2026. Transcript: Adam Stacoviak You create a website today, what do you think about? What do you design it for? We design it probably for a mobile device first and then also for a desktop. Right. And then maybe you think we also need an API because we need to have programmatic access to our website, but you're, you're human first for sure, and you're probably mobile first. Maybe you do it all at once, but he's saying that in 2026, the web will flip to agent-first design. So the first thing you're going to think of, the number one thing you think of when you start a new website first is how will an agent use this? And then comes everything else. I hope that's not true i can see it being like sometimes the case for certain websites for certain uses but like for all the what the web he calls it the web i'm thinking that's an awful lot Jerod Santo Of change quickly for the web which generally moves somewhat slow some pushback on that i i do concur with this thought but the lens i would shift just slightly because i feel like every It's almost like behind every good man is a good woman that that whole shtick or phrase or whatever it's like behind every great human who's doing great work is or maybe in front of actually Since we're front loading this is in front of every great human doing great work is an agent. So I almost feel like you're helping the human do better by being agent first. So I'm sort of, I'm conflicted there because that's how I'm thinking too, really. I think that if this last year has told us anything, people like agents. Developers like agents. I think people like agents and people are going to start using cloud code who don't even code software. They're going to be coding pros or whatever. That's where my lens is at. I'm thinking like if you're not taking the agent in mind, then you're in the past. Like you have to think about, I'm not sure I would say quite agent first, but definitely agent as well. And almost everything, everything, even like simple CLIs are human first, right? You have to, like, if you have an error from a CLI, like you threw a command out there that you misused a flag, have a proper error, not for just the human, but for the agent that gives them Context. Hey, you meant to do this or this is what that does or here's documentation and you give them context i think that's where my lens is at too so i'm not sure if it's for the web and you know agent First necessarily but you definitely have to be thinking about every new interface with agents for sure hands down i agree with that i don't disagree with agent as well which is what Adam Stacoviak You said i think agent first in 2026 is too fast and also i think it's it's both too fast to call and i think it's too fast to do i think agent as well makes a little sense matt what do you think Mat Ryer I kind of get this for sure um because we did we've been through transitions like this mobile you mentioned that was one of them it used to be web before that you'd be building for desktop And it is about how the users are interacting with it that's always been the most important thing about technology so yeah i could imagine like if it's like booking something or if it's Like a hotel or any of that stuff i feel like total that's all gone, it's going to be presentation of something like, here's what I found for you. Here are some images and things, but you want to ultimately, you'll just say to an agent, like, oh, we're thinking of me and a few friends are thinking of going here. Can you ask them for their availability? One of the Fridays coming up would do for me or check my calendar. It goes off and pings their agents to find out their availability if they're interested in it you know and then that all happens they just get some kind of notification and the agent just Asks them a question like a text or it makes the text look like it's come from you even more worryingly probably um and i think that those kinds of flows i think will happen and they'll have To happen but yeah would this mean you're then not going to have an experience where you are choosing what to present agents are all at the moment you are prompting them you have to ask A question for something what if you don't know what to ask it what if you don't know what it can do those kinds of things I think is where we'll have some kind of other experience we had the Same thing with dashboards because someone said, does the assistant project mean you don't need a dashboard now because you can just ask it the question? But there is something about being able to go and look at something without having to ask for it and just having an easy way to go and find that. And I think dashboards still sit there. I don't think dashboards are going to go away for example but it'll be new things alongside it yeah it's the hub and spoke model you always have a you know an api first and you always have Jerod Santo A client it's like if you live in that api first design world then you you totally get what's what's happening here because if you've always been api first a CLI is a client, the website's A client and, you know, an iOS app is a client, you know, that's pretty easy to sort of grok that direction. So I get that. This kind of conflates to some degree with, you know, Paul Dix. The best way to say this, but while I was reading the great engineering divergence, one thing you mentioned was Amdahl's law, which is the principle is when you speed up part of a system, Your overall speed up is limited by the parts you didn't speed up. So it almost reminds me of that in the fact that if you don't think about an agent and you let, and you don't think about that first mentality, like even Matt's describing here, you don't Enable yourself for the things you're building to be as fast as it can be, because agents are going to be much more faster than we are because they're designed to be their machine. I think if you don't, if you don't start curbing that idea, you're going to be dwarfed by the folks who do and embrace these new systems that have to move faster. And they retool their entire pipeline towards agentic things that move faster. And if you stick in the old way, let's just say, you're going to be slower by nature and less fast. So slow to market, slow to think, slow to experiment, slow to fail, slow to succeed, all the things. ([Time 0:55:30](https://share.snipd.com/snip/9e6ededc-db2f-4d8e-821c-4507f4304f0f)) - **Make Errors And Docs Agent-Friendly** - Design interfaces and CLIs to provide machine-friendly errors, docs, and context so agents can act programmatically. - Add structured hints and links that help both humans and agents resolve failures quickly. Transcript: Jerod Santo So I'm sort of, I'm conflicted there because that's how I'm thinking too, really. I think that if this last year has told us anything, people like agents. Developers like agents. I think people like agents and people are going to start using cloud code who don't even code software. They're going to be coding pros or whatever. That's where my lens is at. I'm thinking like if you're not taking the agent in mind, then you're in the past. Like you have to think about, I'm not sure I would say quite agent first, but definitely agent as well. And almost everything, everything, even like simple CLIs are human first, right? You have to, like, if you have an error from a CLI, like you threw a command out there that you misused a flag, have a proper error, not for just the human, but for the agent that gives them Context. ([Time 0:56:55](https://share.snipd.com/snip/c322e114-afee-437b-8355-59fd5f58cb14)) - **Agents Will Stress Databases And Indexing** - Agent data access patterns and heavy semantic queries will stress existing databases and indexing strategies. - Teams must rethink storage, indexing, and embedding strategies to support fast semantic lookups at scale. Transcript: Jerod Santo One was the database access patterns, breaking things, and then vectors. These two stand out to me. You know, why does it stand out to you, Matt, since you're concurring? We've already seen this. We are hammering our databases now. Mat Ryer The agents can do queries a lot faster than humans can, and it can do more complex queries and stuff. Now we're lucky, this is just the example at Grafana Labs, we're lucky because the teams... Where Matt works. That's where I work. We are hiring. There are teams, the teams are very good the loki team like these are like i've never worked with engineers with that particular kind of speciality because you know they care about storage Uh obviously all the data formats are bespoke that they have to invent you know all the indexing all the kind of complexity that they build to make these systems work really quickly so They they're up for the challenge but we are hammering them. We had the drill down apps last year that we did where the UI is basically DDoSing and we've made that even worse now with Assistant. So they have to adapt and change and they're up for the challenge, but is everyone, are some old data techniques or data techniques even going to change? Are we going to start changing how we store data so that it's more it's ready for agents uh indexed differently vector databases i think as well plays into that big time this is where you Create a vector from some kind of content a chunk of content you put it into some multi-dimensional space and then you can query that very quickly and it's kind of semantic as long as it Semantically means like things that mean the same end up in the same area in this multi-dimensional space so you know that this is roughly what you mean the nearest thing to it yeah unbelievable And that powers that makes cursor does this very well the other um the others have the same kind of thing indexing your code base like that cursor makes it basically very quick if you ask A question about the code base it can answer it extremely quickly by just consulting its index that it's built it doesn't it used to grep all the time and it would just take longer to go And grep everything to learn and then it would end up filling its context window too much now it will use the vector database gets the answers right there very quickly and you really feel That difference and i think vector databases are going to be a massive new concern for 2026 yeah vector databases are interesting i'm just barely scratching the surface thinking about Jerod Santo Some of them but um what do you do whenever like you're you said, the vector space that you operate in, like those embeddings are created by, let's just say an algorithm, maybe even an LLM or a model or something like that. And that model gets superseded, you know, becomes part of your architecture to kind of keep your original embeddings maybe, or maybe the original data set so you can re-embed quickly To get maybe even better embeddings as that vector database gets used and there's performance enhancements there's new technologies what do you think about that yeah i think um so Mat Ryer The vector embeddings are different um yeah so we use like there's some open source ones that you can use there are some other there's also like other models that do it um but they yeah So i could imagine there being new vector technology which uh which means you then want to re-index things potentially i think it's better sure but then but the ultimate the in and out Of it to the llm is just here's here's a here's a sort of search query and it returns just results that match semantically so that interface is probably quite safe to keep but who knows Different innovations could happen one of the problems is like clustering like caching right is it like caching yeah it's like a cache it's an index yeah it's a it's an index that but It's an index that you can you to look it up you're just doing simple geometry like they're quite simple um functions to actually uh find the answers because you do all the work at the time You generate in the embeddings and things that takes a lot of time and process to get but but reading it is very quick yeah but but if you've if you have lots of content the more content you Put in then this this vector space can get crowded that's where you end up with problems where it picks things that aren't relevant. You wouldn't consider it as relevant as other things, but it's all too tightly. Jerod Santo Which is when you would do a re-index or a re-embed. You would re-vectorize. I'm not sure what the terminology is for these because I'm just touching it a little bit here. But when the space gets crowded, you need to figure out a way to give distance and give more meaning. Because the whole point is to create meaning and create similarity in the vector space, but not have to stay there forever. Mat Ryer Yeah. So it might be like you would keep a vector of recent stuff. So you keep an index of the recent stuff, and that's what you're going to search, because in this domain that makes most sense. But maybe you also have an older one that contains past things, and you should do multiple requests into this. ([Time 1:16:56](https://share.snipd.com/snip/0585bd5c-a6af-42e5-9993-906de79eb7b0)) - **Re-Embed And Cache Vectors Strategically** - Re-index or re-embed when vector spaces get crowded and relevance degrades, and keep recent indexes separate for speed. - Store and precompute embeddings to trade upfront cost for fast semantic queries during inference. Transcript: Mat Ryer Yeah. So it might be like you would keep a vector of recent stuff. So you keep an index of the recent stuff, and that's what you're going to search, because in this domain that makes most sense. But maybe you also have an older one that contains past things, and you should do multiple requests into this. What about the database? Jerod Santo What do we do there? Just better indexes, more vCPUs, what do you do? Mat Ryer More RAM, dedicated machines i think we're going to end up storing data differently so that we store it so it's in a format that the llm needs which is going to be just natural language In a lot of cases there'll be some cases where it's that for what we do with the assistant just as another example because it is on my mind a lot is we have a small model fun assistant yes so Hot right now so hot so hot well according to a big ai company it's one of the best implementations they've ever seen which model are you guys using we use claude sunnet 4.5 at the moment Um but we're excited about others when they get more affordable but what we use we use a smaller model to look at the data so um we use a so you know it'll make a query but we can easily fill Up the context window too much so we take the data and give it to a different model and say you know we taught it how to describe a graph. It's got a spike here and then it dips, or it's flat generally, and it describes it in natural language. And it's that that then gets fed back to the main agent. So the main agent, it's a bit like saying, look at that graph for me and tell me what you see. Right. And it might, and the agent, there's one will say. It describes it, yeah. Describe the color green for me. Very spiky. Yeah. It'll just, yeah. It's all right. Jerod Santo It's trees. That is interesting how natural language has become the language of choice across these things. Like even if you, like even context is simple, just like it describes it in words. Yeah. I like it because I can read it. Thank you. Yeah. Mat Ryer Until I can't read it right that's what i mean so like the the apis actually like being being just like text that you can read is quite nice for humans as well we don't really do that we have Tends to have json api or something you know but it's open to interpretation. ([Time 1:21:42](https://share.snipd.com/snip/f2690e6c-68ca-4005-91c2-7f8ad0aa6550))