
Orbと永続スレッド、そしてAmpの次の一手(Changelog Interviews #691)
Orbin' with Amp (Changelog Interviews #691)

Orbと永続スレッド、そしてAmpの次の一手(Changelog Interviews #691)
Changelog Master Feed
要約
Changelog の Adam Stachowiak が、Amp の Quinn Slack と再び対談した回。AmpがSourcegraphから独立した20人の黒字企業になった経緯、クラウド上で動く「orb」、永遠に残るスレッドの意義、人間によるコードレビューの衰退、モデルの重要性低下、そしてオープンソースの共有地(コモンズ)が失われる懸念が語られる。終盤では Adam のtailnet上のGitリポジトリ問題を、Quinn がその場で機能として実装する。
この回のポイント
- ●AmpはSourcegraphから独立し、資金調達なしで黒字の20人共同創業者体制の会社になった。トークンコストを気にしない設計で、エージェントの可能性を早く見極めたと述べている。
- ●orb(他社でいうクラウドエージェント)は、監督なしで並列実行でき、手元PCの性能に縛られない。従来のチケット中心の開発プロセスには収まらないとQuinnは述べる。
- ●スレッドは不変で永続し、人間の意図の記録としてコミットにもURLが付く。意図こそエージェントが最も必要とするものだという考えが示された。
- ●Quinn は従来型の人間によるコードレビューは廃れつつあると見る。代わりに動画・スクリーンショット・性能レポートなど多様なシグナルでの検証を重視し、AmpチームはmainにそのままPushする運用だが、高い信頼関係のチームが前提だと強調した。
- ●モデルの重要性は下がり、利用比率はAnthropic系約8割からOpenAI系約8割へ移り、オープンモデルも実用水準になったという。月20ドルのサブスクも導入した。
- ●GitHubが担ってきた「世界中の人とつながって作る」感覚、オープンソースのコモンズが失われることを懸念し、その解決策はまだ見えていないと語った。
章立て
AmpのSourcegraphからの独立
Amp が20人の黒字独立企業になった経緯と、トークンを制限せず可能性を探った方針が語られる。
Orbとクラウド実行の価値
手元のマシンから解放されるorbの利点と、Adamが自作したAmp Runner、jellyware(柔らかなソフト)の話。
tailnet問題をライブで依頼
Adamのリポジトリがtailnet上にありorbが使えない問題を、Quinnが音声入力でエージェントに依頼する。
永続スレッドと人間の意図
スレッドを不変の成果物として残す思想と、コミットにスレッドURLを付ける理由を議論する。
コードレビューの終焉とmainへのPush
従来のPRレビューが衰退する理由と、Ampチームが動画等で検証してmainへ直接Pushする運用、その前提となる信頼。
モデルの重要性低下と料金体系
利用モデルの変化、サブスクの導入、価格差別と補助金に関する見方。
GitHubとコモンズへの懸念
エディタやGitHubが静かに使われなくなる流れと、オープンソースの共有地を守る課題。
資金調達なしの強みとライブ検証結果
黒字ゆえの自由さ、orbという名前の由来、そしてtailnet対応の実装結果の確認。
解説記事
Changelog の Adam Stachowiak が、Amp の CEO である Quinn Slack を迎えた回である。Amp は Sourcegraph から独立し、20人の共同創業者で運営される黒字企業になった。話題はクラウド上のエージェント「orb」、永続するスレッド、コードレビューの変化、オープンソースの行方まで広がる。最後には Adam が抱える問題を、Quinn がその場で機能として実装してみせた。
独立したAmpと「トークンを絞らない」賭け
Quinn によれば、Amp は Sourcegraph から完全に独立した会社となり、20人の共同創業者が集中して取り組んでいる。資金調達はしておらず、黒字だという。
彼は、昨年は他社が月20ドルの枠にトークンを収めようとガードレールを設けていた一方、Amp はトークンを自由に流す方針を選んだと述べる。そのぶん高価になったが、エージェントを縛らなければ驚くほど遠くまで到達できることに、他社より早く気づけたという。Quinn は、資金調達をしていないことが、探索や方針転換、不要な機能の削除をする自由につながっていると語る。企業向けの営業チームがなく、「orb」のような変わった名前も通せるという。
orbとクラウドで動くエージェント
orb は、他社でいうクラウドエージェントにあたる。監督なしで走り、並列数が手元PCのファン性能に縛られない。Quinn は、これは営業からPM、チケットへという従来の開発プロセスには収まらないと主張する。
Adam はノートPCを閉じても、場所を移動しても、エージェントが止まらない解放感を語った。ただ、彼のGitリポジトリはTailscaleのtailnet内にあるため、orbは起動時にリポジトリを取得できず使えなかった。そこで彼は Go製の「Amp Runner」を自作し、専用マシン上で --no-tui 付きの Amp を複数起動して、プロジェクトごとにスレッドを生成している。Quinn はこれを、用途に合わせて作る使い捨てに近い個人用ソフトの好例と評し、チームの Thorsten による呼び名「jellyware」を紹介した。
永続スレッドと人間の意図
今回の中心的な論点がスレッドである。Amp ではスレッドが不変の記録として残り、過去のものも後から読める。コミット時にはスレッドのURLが付く。Adam は、他のCLI型エージェントではコンテキスト圧縮で言葉が失われたり、ログが人間に読みにくかったりすると指摘した。
Quinn は、エージェントが最も飢えているのは「意図」であり、人間が語ったことも、モデルが生成したトークンも貴重だと述べる。バグを修正するとき、コミットだけでなく一か月前の発言を読めれば、本人に尋ねるのと同等かそれ以上の情報が得られるという。なお学習への利用はしておらず、有効化できない設定ボタンも用意していると説明した。Quinn は、エージェントが人間の意図を保持しなければ、人間の望む仕事はできないとまで言い切る。
人間によるコードレビューの衰退
Quinn は、GitHub のPRで同僚がコードを一行ずつ読むという従来型のレビューが廃れつつあると見る。強制されたレビューが残る組織では、並列化の恩恵が受けられず、成果物が滞留するだけだという。
Amp チーム自身は、変更を Amp 内で確認し、エージェントにもレビューさせ、動画やスクリーンショット、性能レポート、開発サーバーでの動作を見たうえで、ship ボタンで main にプッシュする。コードそのものも、重要なファイルなど一部は見る。SOC 2 も、GitHub でのブロッキングレビューを文字通り義務づけているわけではなく、有効な統制を示せばよいと述べた。
ただし Quinn は、この方法は全チーム向けではないとも強調する。長年の信頼と当事者意識(skin in the game)を持つ小さなチームだから成立するという。今後は、少人数で大きなものを作るチームが増えるというのが彼の見方である。
モデルの重要度低下とコモンズへの不安
Quinn によれば、Amp では約1か月で利用が Anthropic 系モデル約8割から OpenAI 系約8割へ移り、GLM-5.2 などオープンモデルの利用も伸びている。最高価格のモデルを使うことより、頭の中の意図をエージェントに渡す手段と、多数のエージェントを同時に動かす仕組み、そして作らせたものを動かす環境が重要になったと語る。料金は月20ドルのサブスクを用意した。価格差別は航空会社のエコノミーとビジネスの関係に近く、補助金とは違うという考えだ。
一方で、GitHub が作ってきた「世界中の人と一緒に作っている」感覚、つまりオープンソースのコモンズが、各自の閉じた環境に分断されて失われる点を懸念している。解決策はまだ分からないとしつつ、核となる10%ほどのソフトを皆で共有し、各自のエージェントがそれを改造する世界を一つの方向として挙げた。
まとめ
この対談は、エージェント開発の焦点が「どのモデルか」から「意図をどう残し、どこで走らせ、どう検証するか」へ移っているという一つの見方を示している。日本の開発者にとっては、スレッドを成果物として扱う発想や、レビューを動画・性能指標を含む多面的な検証に拡張する考え方が参考になるだろう。ただしレビュー廃止の運用は、高い信頼を前提にした小規模チームの話であり、そのまま一般化できるとは Quinn 自身も述べていない。オープンソースの共有地をどう保つかという問いも、今後の重要な論点として残る。
文字起こし(英語・自動生成)
What's up, friends? Welcome back. This is the ChangeLog. I'm Adam Stachowiak, Editor-in-Chief here at ChangeLog. You know what? An agent, as you know, can write the code. They can review the code. They can push your code to main like our friends at AMP. So where should your agents live and who remembers what you actually asked them to do? Quinn Slack is back. AMP is now its own company. And Quinn has a lot to say about orbs, threads that last forever, and why code review, as we know it, is fading. He explains why the model matters less than it used to and why he worries we could lose the open source comments we all love. And Quinn, he even built a feature request for me live on the show to solve my telnet problem with orbs. A massive thank you to my friends and partners at fly.io. Your agents, they need computers, they need machines, they need sprites, they need volumes. They need Fly, and that's what I use. Check them out, learn more at fly.io.
All right, friends, it's Orban time. This episode is brought to you by our friends at Parallel. I do a ton of research. research for the podcast, research for guests, research for sponsors, research for my own projects. And I'm constantly digging into new technologies, new projects, new companies, new ideas. And I'm leaning on AI agents every single day. But here's the thing. Your agents are only as good as the information they get their hands on, but they can access. And that's why I've been using Parallel. Right now I'm using their search, their extract, and their enrich web tools in my own agent workflows. Search helps my agents find the right information. Extract helps them to actually get the content from those pages and enrich helps fill all the gaps with additional context. And honestly, huge difference. It's just kiss difference.
Instead of me spending hours digging through websites, connecting the dots, pulling together the research. My agents, they do all that legwork and bring it back to me, the information I actually need. It's made preparing for shows, researching sponsors, and of course, exploring ideas so much easier. And that's what I really like about Parallel. They're building web tools specifically for agents, not just search results for humans, but reliable access to the web so your agents can actually get work done for you. If you're building with agents, you should absolutely be thinking about how they access the web for you. Because better information means better results. Give Parallel a try today. They've got a free way to get started. Plus, free monthly API credits. Head to parallel.ai and check it out. Again, parallel.ai.
well friends we're back with quinn slack one of my favorite people on planet earth and i'm not even kidding there uh big fan as you may know quinn of what you've done with amp and obviously source graph but how you dovetailed into where you're at right now the last time you had done the pod there was no amp ink i think there was barely an amp if that was even the case I think AMP was a thing, but less used. I think my last complaint was it's expensive because you never subsidize the model or the pricing to anyone. You sort of was like, hey, this is what it really costs, and we're the state-of-the-art. I think you even said the Porsche of Coney. That was Armin. That was Armin. Okay. I agree with that. I concur with that. Every time I used AMP, I felt like I was working with something state-of-the-art. and I would throw my best and hardest problems at AMP and obviously the Oracle must bless everything.
So that's my introduction to you and AMP and everything else. What's going on? Where are we at? Well, thank you. It's awesome to be back. There is so much happening. I mean, just from the way that I am building now, it feels like it's changed more in the last eight weeks than it changed all of last year. and yeah last time I came on I was at Sourcegraph and now fast forward to here we have spun out AMP into its own company so Sourcegraph and AMP used to be in the same company now AMP is its own fully independent company that is a very interesting kind of transaction to do but it was absolutely the right thing because now we can focus and we've got 20 people on AMP we set out with 20 people and we said, we are going to just get super close. We've got 20 co-founders. We are profitable. And we want to explore where all of this goes with AMP. And last year, that meant not worrying
about token costs. You know, everyone else was trying to fit in all of a person's tokens into $20. And they were saying, I'm going to give you guardrails for this and this and this and all of that. And the model was the most important thing there. That was last year. And we broke that trend in a few ways. We said, well, let's explore what happens with AMP if you actually just let the tokens flow. And it made it more expensive for sure. But it got us to see something earlier than a lot of other people building agents, just how far they can go and how if you don't keep them on a tight leash, they can do incredible things. And now, you know, everyone has seen that now. And it feels like that for all the change last year, going from a human, you know, before agents, maybe using chat with AI to ask questions, and then the agents started to write code for them and take multiple steps for them.
That obviously was a huge leap. It felt like that let us all get more done, but in the confines of the existing way that we were working. We were still, not at AMP, but for most other people, still using the same process where if they're building software, the salesperson would come talk to the product manager, would come and make some nice linear GitHub tickets or JIRA tickets, and then you'd now build them a lot faster. But now with Orbs, and other people call them cloud agents, the ability to run agents on machines where you don't need to supervise them and where the parallelism is not limited by how good your computer's fan is, that totally changes how you work. It means you can do so much more, and it can't fit in this old process. Yeah, I've become spoiled to agents not running on the machine I'm actually on. It's liberating, you know, to be able to close the lid to your laptop
or to not be anxious if you've got to transition from one workspace to another, like physically. I've got to go from this room to maybe in my truck with my kid to his ninja slash parkour class, and there's a working space there for everyone. So if I've got to go between internets, what happens to my agents? You know, like this anxiety of... Yeah. So what was it like for you to switch from a single agent on your machine to living in the cloud and orbs? I'm not quite there yet, but close. The way I liberated myself, I would say, and my agents is I gave it a home. And my challenge with AMP is that my source code lives on a Git repository that's only accessible via the tail scale or via my tail net. So I've never been able to fully enjoy orbs the way orbs allow you to be, which is sad, but I've used them. And what I've ended up having to do is create this system where my agents pack the code up, throw it onto an orb,
and still it has an issue because of the boot up because it has to pull to get the repo. So we sort of fake it. But the way I was able to liberate myself and some of the things I'm working on was to put it on a dedicated machine that was Amplified. and when I say Amplified, I'm using the amp-dash-no-tui and I built a Go, let's just say an orchestrator, an amp orchestrator I call Amp Runner. And so I have something called Amp Runner written in Go that orchestrates all my projects across this one machine all spinning up with dash-dash-no-tui, which means you know how this works, but audience, there's no-tui. Like it doesn't launch the amp-tui And so the way you access it is via the web, via the URL, and that's been a pretty phenomenal experience. Like to have a threaded experience versus a this is in a CLI bound to this machine. Sure, it's still bound to that machine, but I'm not stuck with that terminal experience,
which isn't bad, but I'm not stuck there. Now I can move around. I've got the web at my hands, so I just go to antcode.com, and this is not an ad, but I'm a big fan, as you probably know, enough to build an orchestrator. And so I thought it was pretty cool to be able to orchestrate AMP beyond even your levels. I'm leveraging your API. I'm leveraging your CLI. I'm doing all sorts of things. I've even got, like, messaging behind the scenes above and beyond your threads that let them go beyond a machine. It's pretty wild. Wow. I had no idea you were doing all of that. That's incredible. Yeah. I don't publicize it because I'm a closet technologist, let's just say, for now. I'm just so focused, so focused. I feel like so many people out there are just like, let me find a way to build something to showcase what it is, and I'm just too stuck and too fixated on building something, and I just can't stop building it. So every time I want to broadcast Amprunner, for example, it's like, well, that's not the point. I'm not trying to – I'm trying to make it for me for now.
Yeah. You know? Well. But at some point, I might. Yeah, I'm glad that works for you. And that actually gets to, I think, some of what I am seeing in how I use software differently and run software differently, where it is so much easier to build kind of custom, temporary software just for yourself. Yeah. And this is not, you know, like last year it was, oh, this company Vibe coded Salesforce from scratch. And, like, that stuff, that's dumb. What you did is not that. One, you're not going out and trying to hype it up. And two, it's something that actually works for your case. And you're saying you solved a problem for yourself. You're not saying you went and cloned some other thing, because the important thing is it's about solving problems for you. And this promise of personal software or something softer than software, that is... Softer than software. I like that.
Yeah. Thorsten on our team calls it jellyware. Jellyware, okay. I kind of like that. I thought, you know, you spread jelly on toast, and then it turns out all the Europeans on our team were thinking it's like Jell-O because I guess, you know, we call them different things. Right. But either way, softer than soft. So the alternate experience would be for me to SSH into that machine, which is, you know, not this MacBook Pro, for example. It's a different machine on what I call a dev lab now, not a home lab. I'm a dev labber now, not a home labber. It's slightly above the home lab echelon, just so you know. Yeah, a little more refined. Agents just don't take it serious when you see I'm in my home lab. They're like, oh, it's just this way, you know, this and that. I'm like, no, this is a shrimp laboratory, okay? Like, there's several Proxmox knows there's 35 VMs on this network. They're all doing cool things. I mean, you've got lights. It looks pretty badass. Like, come on now. Let me take – not that home labbing is bad, but it was like just keeping me –
like the agent would never fully understand the level I was trying to operate at. And it wasn't a pejorative. It was just like, well, this is just a toy in the home lab, not a developer lab, building a developer tool and all sorts of fun stuff like that. So the experience with AMP in particular using the dash dash no 2E flag on your CLI would have been I would have to go to every directory that would instantiate a new agent to run there at a thread, right? And so I didn't have to do that. I built an amp runner, which would orchestrate a brand-new directory for a new team on that. It's a brand-new Git checkout. It's not a work tree, and it would be its own thread. And because threads can talk, they can talk, which I think was so cool. But I would have to go to every directory and manually do that. Well, amp runner just does that. I'm like, hey, I need a new project. And so then I actually created one system agent that was ran by AMP on the system. And then that system, that thing there got the most privilege. It got API tokens and secrets that none of the other agents would get, and it could orchestrate this system.
And as I'm building out different projects or different teams, it would instantiate them using AMP runner and just spin up dash dash no to it. And next thing you know, there's not only a session with Dash Dash Note 2, but a thread that it creates to my repository. It spins up all these things. It was orchestration for me. It was around AMP, but largely built and predicated on how I built software, how I built Teams. Yeah. Yeah, that's incredible. I mean, basically you built this orchestrator that you made your own orbs, kind of. You forged the raw metal of the machines in your dev lab. You're damn right. And you made your own orbs. That is awesome. Hence why I wasn't so upset. Now, I love orbs. I love the concept. I haven't personally. I do understand what they do. I understand how I will use them. But the only limitation, Quinn, please, as we speak on this podcast, in your other screen, vibe code quickly. I mean, it's a bit of a door to, I don't know. You know, hop in AMP and say, Adam has a problem. please let me instantiate an orb with a secret that is a tailscale API key
before the Git repository is required for the orb to be instantiated. Because that's my only limitation is using orbs. It's like I can't use them because my Git repository is on a tail net, not GitHub. All right. Let me share my screen, and I'll kick that off. Okay. Let's put something. Sounds good? Okay. Here, I'm going to open up AMP over here. I got AMP open, and I'm going to start a new thread. I will run it in an orb, and I'll say, I don't know what I was doing before, but that can come back later. These drop-downs below, can you narrate some of these drop-downs below just so folks can track? It's in an orb. Maybe not so the directory it's in, but maybe how you're choosing your level, even like you're on medium, which is my favorite, by the way. Yeah, that's right. So I'm in the new thread screen in AMP, and I'm going to do it in an orb. I usually have my Mac machine set up as a runner, but I just rebooted it. So I don't have that there, but I could run it on my Mac machine.
I do not have a DevLab. But your Mac machine is a DevLab, basically. Yeah, yeah, you could call it that. We'll get you amp-runner set up in flurry. And then projects. We've got a monorepo, so I'm going to run it in the AMP project. I like monorepos, man. Monorepos for life. Makes it easy. And then the cool thing, now if you're listening, you're not going to see this, but I clicked the mode, and now I'm choosing the amp mode. And it looks like one of those old-time stereos, the professional audio equipment where you can go low, medium, high, ultra. Ultra, of course, goes to 11 if you count the notches around. And we've got these four modes that are built in, and we curate these. So usually medium is a really good default. Ultra is when you are building something super complex or when nothing else has been able to do it. And you can see exactly what each of these means. The ultra is Fable 5 as the main agent,
and then with an Oracle of GPT-5.6. But I usually stick with medium for most things. That's GPT-5.6 sole. You can also go and pick any other model. We've got a lot of other models. You can add a bunch more. And we'll stick with medium for this, though. So I'm also super hooked on dictation because the fastest way to get stuff from your brain to the agent, which is now the bottleneck, that transfer of information is dictation. So you ready for me to dictate? I can't wait to hear it. Let's do it. All right. I will say I'm talking with Adam from the changelog now, and his repo is on a tail net via tail scale, and amp orbs do not work when that's the case. We need a way at the very beginning of cloning or even as part of the repository configuration to make it so that we run something so we can get onto his tail net,
and it can use like an environment variable or whatever that he's set up. but we need something to make it so you can clone a Git repository that's only accessible on a tail net. So I was dictating to the agent there. Now it got the text, and it's kicking off a new orb. I love that. And it will be added. Now you can see we've started this thread in AMP. It's going to go and work on that. And you can see all the other threads that I have going on. we've seen that as the number of things you're doing increases it yeah gets a little bit crazy so the vertical tab is a good idea versus the horizontal tabs for for the amp interface too because obviously infinity yes yeah but we got to figure something out it's getting pretty unmanageable so it's going and we'll let it go we can get back to it later and check in on it But I will, you have basically made your own orbs and orbs, other people call them cloud
agents. Yeah. We call them orbs because a lot of people tried cloud agents in the past or other people's cloud agents and they just didn't work as well. And they got turned off from it. We made them really good and we wanted to come up with a different name because we kind of like being different, being weird, being fun. but also it gets a lot of people that had a bad experience with cloud agents and other products to give it another try and to see if all this stuff is actually integrated and works really well, then what does the experience feel like? So it sounds like you've got that good experience, and I would hereby say you can call them orbs. Sweet. All right. But we can get you our orbs as well. Yeah, I look forward to that. But, you know, the thing I think was what I was trying to do, especially when I saw this dash dash no two flag and I understood what it did. I started digging into the manual slash docs. You call it the manual, which I think is cool. Again, being different.
You know, you don't give somebody a set of docs. You give them a manual. This is how you use a thing. And that thing is so detailed. It really is. And so I went even as far as like because every manual URL, as you know, it ends in .md. so every URL you have is a markdown file. I built an ETL pipeline to consume that, and as it changed, my agents would pay attention to new. So I get your updates via, like, everybody else, you know, Twitter and stuff like that, or X, but my agents also consume it immediately through your manual. And just because I'm such a fan, I like AMP a lot. That's awesome. But I think, you know, my next move was to do something like instantiate a new VM or a new container with AMP installed and running dash dash no to it. Not because I want to go around your orb world because I couldn't get in. I couldn't get in. So like any hack, I'm like, by any means necessary, right? And as a fellow home labber, dev labber, you know, I've got too much compute to go around basically and VMs out my ears.
So I not trying to you know pejorative an orb but it a VM It a VM in the cloud that you run on probably AWS or wherever you can get your compute Me it localized right here on this machine I can spin up a new VM pretty fast with Proxmox A lot of that is orchestrated and automated these days. And to put AMP on there with an API key in dash dash node 2 is pretty freaking easy. So I hadn't gotten there yet because my machine as it was was just fine. I really didn't need to have more VMs and more I'd already had that compute separation between my laptop and a dedicated machine. And then obviously the way that you orchestrate URLs and threads, I think the primitives that you built around, like the thread primitive, I think you all chose some really interesting engineering tactics to give yourself these core primitives to walk around because the moment you have a session, you don't actually have a thread in the case of dash dash notui. you have to supply the thread
which is either go now to the web URL and click buttons like you just did but I didn't do that right my amp runner's like no no no Adam you can't click buttons I will do it all for you and it happily gave me a thread and it's like boom here you go and so I go now to ampcode.com and I look there and if I have a new team or a new project it's just there a new thread is just there I love how you're like if you need to retire a thread or start a new thread, you can consume or take over an old thread. That was built into Handoff even. I think that may even be like the glorified version of what Handoff has become because you used to have a context window. It used to get totally consumed, and then AMP would yell at you and say, hey, listen, now you're like 87%. You've got to do something about this. And now I don't even worry about context because you all figure that out. But I'm rambling about AMP. It's pretty wild to liberate yourself from, you know, If you're on a laptop right now and all your agents are running there and you're anxious, there's a good reason for it. It doesn't have to be that way.
There's a different way. There's something better. There is something better. That's right. We want to get these primitives right. That, I think, is the most important thing because this is a golden age of computing. We're all figuring out how are people going to be using these computers? What are the building blocks here? If you think back to when the Internet was being created, things like the Internet Protocol, TCP, UDP, all of that, these were ideas that people had as building blocks. When Unix was being created, the idea of processes and files and what a file is, there were so many other ways that you could have thought about a file. You had other operating systems like macOS Classic and BOS and even Windows having different concepts. and getting these concepts right is so important. So threads, I think, is one that we got right. And what's really cool is if you've been using AMP, you know, the whole time or since when you started, those threads, you can still go back and look at them.
Yeah. To pause there for a second, like that core tenet right there alone is the icing on the ant cake, so to speak. To be able to go back to a thread from seven months ago, and it's immutable, it is what it is, but to be able to recall it or even to mine it for ideas. Like how often do we pour via voice or via keystrokes, we pour a massive idea via a prompt into our agent, and it's sort of like ether. It just sort of like just goes by because it's a linear thread. But to be able to recall those, and for that to be the way the – do you call it a harness? Do you call it an orchestra? What do you call – what exactly is Amp these days? Like, what do you call that? We call it an agent right now. Okay. But, you know, it's a product. Right. It's something that humans use. The reason I ask that question is, like, you know, you've got other CLIs.
Like, you can go to Codex directly, right? Codex, dash, dash, live dangerously, whatever you want to spin up in there. You can do the same thing with Cloud or any other CLI-based agentic workflow. And AMP is very much like that too. The difference is that I worry. When I pour my heart into a Codex TUI or into a Cloud TUI, I think like, gosh, they're going to compact my context. My words will get jumbled. It's lost. They don't give it to you as a true artifact as a developer. It's like taking the scratch paper you write your ideas on and burning it immediately. Although it is there, it's in a JSON-L file, but it's not a human-readable file. They don't really give you tools to spelunk it. Whereas AMP has thought of that, like from the beginning, this thread has to be important enough to be able to recall it a year from now, two years from now, whatever the number is. I mean, that to me is a really interesting primitive. Other tools, they don't expose this. Sometimes, like with Cloud Code, it'll even just start on your local machine and wipe it after a certain amount of time.
And, yeah, it's exactly like. It makes no sense. Like, why would you do that to me? Yeah. It's not for me. Every single thought that a human has had about what you're building is precious. And also, every single token that a model has produced is precious. And we set out with the goal of never wanting to lose any of the human ingenuity and thought. and we never want to lose any of the inference either. Yeah. And we store that for you. We do not train on it, to be clear. We're not storing this for our commercial purposes. We have a button where you can enable training. We added it there because a lot of people wondered, do we train? And that button is a no-op. And for basically everyone, it says you are unable to enable training. So we even had to put a button. I didn't even notice that. That's cool. Yeah, it's really hidden. But there's people who go there. We do this because the intent, what you're trying to do, is so important.
And that is the thing that agents are starved for. That's why I use dictation to get the thoughts and desires from my brain into the information to the agent. That's why we make it so that AMP can access all of the historical threads that you've done with all of your human inputs. That's why we make it so that it can read other threads. That's why we make it by default so that when it commits, it adds a little URL to the thread. Because if all the agent can go off of is the commits, that is usually not as good as if it can see what the human intent was. because if a human said something a month ago about, I want this, and then there's a commit, and that commit had a bug, it's hard to know what the human's actual intent was. But knowing that, I mean, that's exactly how you would build. If you were going fixing that bug and you didn't know what the person was trying to do, you would go ask that person. And by giving the agent today the ability to go and see what that person said a month ago,
that's just as good in many cases, better often, than you going and asking that person. So it's about respecting that human ingenuity and intent that went into building. There's no world in which agents can do well if they do not preserve and use the human intent. And I have chosen my words carefully there. When I say do well, I mean there's no way that agents can do what we want if we don't tell them what we want. And threads are just about getting it to know what you want. And that means over long term, not just today, not just for your prompt today, but really what you want over the long term. And, yes, it's like the other agents sometimes that feels like they're destroying the scraps. It's like you got George Lucas and he made Star Wars, and then you're going to destroy all of the notebooks, all of the character development. I mean, what a shame. You would not do that. The film's there. Go watch the film, right? Yeah, exactly. Don't leave the notebooks. This episode is brought to you by WorkOS.
Agents aren't just writing code anymore. They're accessing customer data. They're taking actions across the tools we use and depend on. So who decides what they're allowed to do? That's a big theme from WorkOS and NIT2026, their conference for builders. They preview three products that show where things are heading. First, Airlock, intent-based access control for AI agents. In a live demo, an agent worked across Linear, Snowflake, Stripe, and Gmail. Airlock denied a large refund, sent a smaller one to a human for approval, and blocked an email containing financial data. And then there's Atlas, a secure home for internal apps built with AI. Think authentication, data connections, credentials, permissions, and deployment. The stuff a prototype needs before it's safe to use at work. And then they have Passport. WorkOS is pushed towards zero sign-in. Enroll your Mac once, then open WorkApps already signed in. All three were announced as previews, but the direction is clear.
As AI makes software easier to build, we need to get more serious about who and what can access it. I'm linking up the Init 2026 recap, so check it out at WorkOS.com. Link is in the show notes. Thanks again to our friends at WorkOS for sponsoring ChangeLog. Once again, WorkOS.com. You know, there's a lot of ideas, though, that get lost in the ether that don't actually hit the mat to become product. You know, like an idea that's big and visionary, let's just say, that requires, you know, a certain roadmap to get there. But you get the point A and point B, but you never get the point C, D, and E. And the C, D, and E is stuck in that thread. and if you can't even for yourself remember because you're moving so fast, what did I talk about during XYZ? Or as you mentioned, this bug, you know, if the pull request, and I don't know if that's going to go away or not because there's still a lot of debate if pull requests go away or not, even if there's a new Git host.
I know you all have a Git hosting now. I'm not sure how robust it is, but I do know you have that. To have that thread become part of the artifact lineage, let's just say. You know, if there's a bug or even a feature, the feature pull request has the thread ID, potentially even the thread URL attached to it. So if ever there's cause for a reason to go back to root cause or spelunk or figure out what was the intent here, and maybe there's a gap that was missed, we have the thread of what the original intent, as you said before, was. So now you can be like, okay, now I have almost better than root cause. You have the real conversation, the real narrative, and the agent forgot to do X, Y, and Z. And the human, because they're moving fast, they didn't notice it because that happens a lot in today's non-code-reviewed, pull-requested world. That to me was just like we're camping on threads deeply, but that core primitive alone feels like – and then the immutable factor and the never-go-away factor is like the true building block that made AMP become what it has to become.
because you have tenants, core tenants that it can't really navigate away from. The fact that threads are always there, well, that's true. It's forever true now. Now I can operate differently. As a human that's building software fast, a heck of a lot less anxiety about where my code's running and if my words are captured forever or being thrown away or burned. Like, no, don't do that. Yeah, yeah, that's right. And you mentioned code review. I think that's another interesting thing. Code review, you can think of it as just making sure what the agent did is correct, and that is always important. But if you think about code review as making a change, going and creating a GitHub pull request, having one of your colleagues usually go on the web and read through the code manually, read everything, and then go back and forth, that, we're seeing that die. Now, at different rates in different places, and it's that kind of change in the process, the human process that goes on inside Teams.
That's the biggest reason why people are having totally different experiences of what it's like to work with agents, because if you still are required by your company or if you still feel that it's the right thing to do to have that old human code review process that's mandatory, then you're not actually going to benefit a lot from the ability to parallelize your work. It's just going to mean more stuff is piling up. And if it's going to take a day anyway for a teammate to review it, you actually don't really care about making something go from 30 minutes of friction to zero minutes of friction. So I think, though, that this idea of code review, it's got to change if you are to benefit from what agents can do. And we're not saying don't review code anymore. Actually, on the AMP team, we obviously are very heavy users of AMP. And it's incredibly rare that we will YOLO push something, that we'll do the, like, Google, I'm feeling lucky,
where we just have a prompt and we just send it out. And we push it to main without ever looking at the code or anything. That is not something that we do on AMP. We'll do that on personal throwaway software that we're building. And the only time we would do that on AMP itself is if we're making a change that we know is going to be only to some site admin functionality that we have or a typo fix. But what we do look at is all these other signals the agent can give us. So if you think code review more expansively, you're reviewing the output. You want to see the agent showing you a video. What is it like? You want to know that the agent, in some cases if it's in the UI, has inspected every single frame of the video to make sure that there's no jitter during the loading of something. For example, you want to see screenshots of how it looks. You want to see the performance report from after it's run through that. And that, and maybe you then look at 10% of the code.
You look at some key files. You inspect. And let's be honest, it's not like with pull request code review. your colleagues were looking at every single line of code. And in most cases, they were not even running the code. So, you know, with AMP, you also have these portals where you can run something and see your dev server, see what it's actually like, and it just works. And, you know, computers are powerful, but it's a huge pain to have like five different dev servers running. So we give you all these other ways to verify, and we're trying to get you to the point where you can, in as little time as possible, have the right signals to know that something is good. And I know that there are ways other than your colleagues looking at all the code for you to get comfortable with it being good because the thought experiment is this. It's, you know, what if AMP had a billion-dollar guarantee that if we got anything wrong in your code, we would pay you a billion dollars.
Now, don't get too excited. Let's go. Let's go. Yeah. Maybe we'll get there at some point. But in that case, well, then you say, yeah, I would click that button happily. So once we've established that there are alternatives to humans reviewing code, then it's just a question of, okay, well, not a billion dollars, but how do we get you to that same level of certainty? What if you saw a video? Yeah. And you've got to wean yourself from that. And a lot of people say, well, but what about policy compliance and so on? And things like SOC 2, they do not literally mandate that you must have blocking code review on GitHub. What they require is a set of controls that you've shown to be effective, and you need to follow those. And so actually one of the most viewed posts that we've done in the last couple weeks is when we talked about how we are able to push to main on a product that's used by a ton of big banks and other companies, and we also have SOC 2.
Yeah. I was always curious about how you actually accomplished, you know, I imagine you still branch. Very rarely. How do you push the main? Do you just never branch? How do you do that? Explain. Give me a one-minute version of what you mean by that. When we're building, we'll go make a new thread in AMP, and that doesn't usually create a new branch. We'll look at the changes there. In some cases in the thread, it's going to make multiple commits. We can review it all in AMP and then we push it. And there is a way in AMP for you to say, instead of pushing a main, go and when you hold down that shift button, it got a nice animation, it'll push you a branch or PR for you. So you can configure it to do that. But no, by default, the work is just usually uncommitted. We review it within AMP, and we have the agent review it as well. And then when it looks good, we've tried it out. We've looked at videos. We've played around with the dev server. Then we just instruct the agent to push domain with that ship button.
Wow. And once you go there, you cannot go back. If you have experience in your team, though, these are not brand-new developers who are, you know, born in AI. They may have actually written lines that code themselves before them, right? That's right. Experience and accountability. and we're not saying that every team is ready for this yet. With AMP, we have chosen a few things so that we are able to move faster, experience things before so that we can get out ahead and we can write about this. We can build a product to hopefully help people manage the changes that are coming. But with the AMP team, as I said, when we spun out of Sourcegraph, we brought 20 people over. And these 20 people, we have all 20 people We've hired one person on it, but otherwise it's the same gang. And we all have worked together for a long time. In some cases, eight years working together. We all trust each other. Everyone has a lot of skin in the game.
Everyone is a co-founder. Everyone has a big stake in the company. And that level of accountability and trust, I've worked on great teams before, but I've never felt anything like the level of trust and skin in the game that people have here. It's truly an incredible feeling. And so, yes, what we do will not work if you have someone who just joined and is very new to software development, or if you have a team where the trust level is not as high on average. And there's a lot of good reasons for that. When I say that, it's not like, oh, you have a bunch of reasons why people shouldn't trust you. But no, if you have doubled your team in the last four months, or if it's people that you haven't known for a year or more, it's hard to get that level of trust high enough. But our belief is more and more teams will come to look like the AMP team, because with agents, you don't need as many people to do things.
And I think that overall, the world needs so many more people building. So this is not a dark and gloomy view about jobs or anything. It's a view about the average size of the groups that build incredible things. I think that the size is going down, and that means that on average, everyone is going to have more of an influence on the product. It's going to have more of an ownership stake, more skin in the game, because if you have that configuration, then you're able to do so much more with agents. And what I love is that makes the world more creative. There's nothing that strangles creativity like being a cog in a big company, like taking tickets from one person that has filtered them from five other people, and then once you're done, you're a tiny little part in the process, and then you – that is not where true incredible technology, disruptive innovation, all that, whatever words you want to say, that's not where that gets created. It gets created from small groups of crazy people building stuff, And that is how I think more and more software is going to be built.
Yes. Yeah. And that's been my lens of AMP since the beginning. Like, I can barely recall the exact homepage originally. And I think Torsten tried to take the credit. I think it may have actually been your words. Bjorn may have said it was your words Bjorn was your co in Sourcegraph I think it was Thorsten Okay The original words on the ampcode website I can recall them exactly, but they were, it was a moving set of words. And you could tell that the ambition behind the team, behind AMP, had it wasn't like a straight and narrow road to Proct. It was like, let's really explore. Let's push the boundaries. I didn't know that you were just like, let's just figure out what happens when there are no tokens or there's really no strangle on tokens. I didn't really understand that to be the case. I knew it was expensive because of those reasons, but I never knew that was the
case behind it though. So you've all been able to build something really interesting, but you've come so far beyond what I think would be what Armin called the Porsche of agents to now you have get hosting, I think. You've got secrets and variables that you can stamp and plug into projects. You've got all these different primitives. What is the world shaping into for AMP? What is the future for AMP? Yeah, the stuff around us is changing. In the past where it was so much about the model and using the single most expensive model, now that's totally changed. Nikolai and our team wrote a post about a month ago when we switched to the dial, that thing that's got like low, medium, high, ultra, that is a single linear scale, basically. And most people don't even know or care what model is being used. And we went about a little over a month ago from about 80% use of Anthropic models,
Fable and Opus in the product, to about 80% use of OpenAI models. and our customers' use of GLM-5.2 is growing a lot. The model, they're all pretty good, and open models are catching up. They are really good, too. Like, they're actually good. People have been saying they're good forever, but no, now they're actually good. Okay, they're actually good. GLM-5.2, yeah. And the model doesn't matter as much anymore, and what that means is you can have an incredible outcome even if you're not using the single most expensive model. And that changed what we wanted to build with AMP. So we introduced a subscription. We let you use your ChatGPT subscription, your SpaceX AI subscription. We've got much cheaper models. And soon we're going to ship support for anything you want to throw at it, like open router and open AI compatible endpoint, all these other things. Because now what is most important is the thing that gets that information out of your brain
to the agent and then the thing that lets you run a bunch of agents at once. And we're not stopping there. I think what's also really important is how then you can run and use the things that agents build. And I'm not talking about the awesome stuff that you built, Adam, which is like kind of a lower level. It's, you know, you can use AMP to build that. But it's like the software that solves problems for you. And that is, I think, a new frontier. Now that this promise of personal software can actually be realized, you know, not like last year with the Vibe coding Salesforce, but we're seeing this actually be a thing. That is where the interesting frontier is. It's no longer about using the best model. The models are going to keep getting better and better. But in the past, remember, everyone used to switch to the new model on launch day, and everyone was so excited about it. And now it's still exciting to see, but you don't really trust the benchmarks. You often don't even get access to the models on launch day because of government restrictions on freedom of intelligence and all those things.
And so it's not about the models as much anymore. That's been a huge change. So now, I like to think that we still have the Porsche quality and experience, whatever. I drive a Subaru and a Tesla, so I'm not a big car guy. But you can use AMP for $20 a month. That's our subscription, and $10 if you're a student. Which I think is phenomenal. You got megawatt and gigawatt as subscription levels, which it seems the main differentiator is the amount of credit you get, which was hard for me to understand. Maybe you can explain that or maybe not. And then just like it seemed like almost unlimited orb usage, like 1,000, I think. Some dramatic number from your $20 plan to your $200 plan was pretty tremendous. I'm excited about that because now what to me has been like the creme de la creme of an agent to use or an agent harness or an agent framework or whatever you want to call what AMP is now, it's always been less than accessible because of the finance for obvious reasons.
Now, Sourcegraph has a ton of enterprise customers, so maybe that was easier for you to get a lot of those folks to be like, let's just try AMP because they've got deep pockets or budgets. but for any developers paying for their own tokens, you've got to be a bit more cautious with how you spend your spend. And subscriptions are nice in that case because it does give you a surplus. It does sort of hide the actual cost of tokens because if ever you've had to pay for overage costs, you're like, well, dang, that subscription is great, but I spent just as much, if not more, in my subscription cost for a fraction of the tokens, right? You have just a literal fraction of the token. So they're giving you generously a lot, which I'm not sure how will play out in the grand scheme of things. But nonetheless, I think all that will eventually figure itself out. It's here to stay. Yeah. I think of it like an airplane. There's the economy seats and there's the business class seats. Yeah. And people that are saying, oh, the token subsidies are going to go away and subscriptions are going to stop being so generous.
It's as though they're saying United Airlines, they're going to stop having economy seats. and they're always going to change the mix around. They can always change pricing, but when you have a really big upfront expense to go train a model, then if you look at any of the economics theories around here, your optimal way of distributing that and selling that is to essentially sell it to anyone who will pay above your marginal cost of inference and then you layer on that wanting to lock out your competitors In some cases, these companies are also gathering data from some of their cheaper users. And all of that means that they're always, for them to maximize profits, going to want to price discriminate. And that sounds like a bad word, but that's basically just match what a big enterprise is willing to pay, which is a lot more, make them pay a lot, and make people that are doing this in their dev lab at home or who cannot pay that much, they want those people still to pay, but not to pay as much.
So that is here to stay, and I don't think there's any world in which everyone has to go pay for tokens or the subsidies dry up because fundamentally I don't think there's subsidies in the same way that it's not like United Airlines having economy seats is a subsidy. They're making money on those seats. Yeah, that's true. How has, I mean, I've got you on record at least at one point in your life on this podcast and you don't write code anymore or a heck of a lot less or you don't write production code, which I think is probably not true today. How has your journey as a CEO of a company and then spinning out a separate company, which was a product for a little bit and then an actual company, which is the case of Amp Inc., how has your journey as a software developer changed from not writing much code and certainly no production code, which would be the pre-agentic era response from a responsible CEO, right, to now? What is your response to that question now? How has it changed?
Right. When I was CEO of Sourcegraph, our product was code search and still is and a lot more relevant now than even a few years ago with the growth in code, which is awesome. I've got the SourceGraph board meeting after this, and, you know, the company's doing really well. The focus that we each get is very good. But, yeah, you know, 200 people on the team. I was CEO. I was responsible for all of that. And I love code. That's not already clear. But, yeah, it was not a good idea, in most cases, for me to write production code. And it really pained me, though. And what pained me is, you know, especially as I saw AI getting better and better, I knew that it was going to change, and I hated being far away from the product and customer. I was very close to the customers in that I was meeting with them all the time.
But sometimes I wish I could just go fix the bug that they said or go build the feature that they said. And there was an apparatus, a process. And that's what every company had at the time where, well, let's put that in the roadmap. Let's educate marketing and sales and PMs and developers and why this is needed. And there's so much work that went into making the case for why this was good because the build part was important. So, yes, I guess before some other people, you know, I was not writing production code, but it was because I was a CEO. And obviously I love getting back into it. And with AMP, I write a ton of code through the agents exclusively. It's very rare that any of us on the team are getting in and actually writing code in the editor. I switched to Zed from VS Code maybe like four months ago. Zed is awesome. I think it is the best editor, but I just don't use the editor that often.
This episode is brought to you by our friends at DeepGram. Their speech models are built to understand words in context because context really matters, as you may know. Like if someone says there's a compatibility issue, they don't know if they're talking about technical difficulties due to conflicting interfaces. Or if they mean you want kids, I want a boat, this relationship is never going to work. Using context, DeepGram can tell one from the other, making conversations better. I use DeepGram. I love DeepGram. I use it for all my transcripts now. It is amazing. It's my intelligence packet. And DeepGram's speech models and real-time infrastructure enables AI to keep up with the way people actually talk. Over 5,000 organizations globally, me included, use DeepGram's speech-to-text, text-to-speech, and voice agent APIs in production. Built for accuracy, low latency, and cost efficiency, DeepGram makes high-quality voice experiments practical to build and operate at scale.
Available through cloud APIs or self-hosted and on-premises deployments. Voice is the most natural interface, and DeepGram is bringing it to the technology we use every single day. Again, I'm using it. I love it. The future is talking, talking with Deepgram, and now when you spend $1 on Flux, text-to-speech, you get $1 back, up to $500. To sign up for free or find out more, go to deepgram.com slash context is everything. Once again, deepgram.com slash context is everything. Yeah, I'd kind of let go of the editor, too, in a lot of cases. is similar. I would keep Zed has always been my favorite editor. Never really cared for VS Code. No disrespect to the team by any means, but it just was not a tool for me. It was a tool for everybody else, basically. Dramatically slow in some cases, just focus on the wrong future, basically,
in the last year and a half. Yeah. And Zed being a very fast, high frame rate editor alone, super fast, written in Rust, great team. Nathan Sobo and team, they're constantly innovating over there, which I'm just so impressed by that team and what they've done too. But the editor itself was great, and they put a great agent experience in it. It just never fit for me personally because I had different areas I'm trying to go into it, but I applaud them for trying to bring the agents into this very awesome editor. Now I just sort of just trust the model in some cases. Obviously, trust but verify. I still look at pull requests I'm not examining them to the nth degree but I'm sort of like does this give me any icky feeling it's kind of like how I look at it now I'm not deeply examining it but I'm certainly going to check it a little bit and it's only because you should be looking at the live app and being able to play around with it you should look at videos from it
and sometimes those will be attached to the PR but sometimes they're not and then if you're looking at the PR UI It's like a dead interface. You can't say, give me a video of this. Or you can't, if you've got a bot rigged up, you can ask. But it feels kind of clunky, you know? It does. It does. But, you know, with editors, I think when people go and say, I don't write code in my editor anymore, there's a lot of people that nod to that. And then there's a lot of people that say, what? That's crazy. And I think that those people that think it's crazy, they believe that the person who switched from their editor, like, took a week and said, I'm going to get off my editor. And it was, like, a really hard thing, and it felt like a big change. But no, I mean, actually, all these things, they kind of die with a whimper. They don't die with a bang, like with GitHub. I was user 1976 on GitHub. I guess, you know, Toby from Shopify, I think, said he had a museum-grade GitHub ID that's much lower than mine. But I've been using GitHub forever.
And now we still push our code to GitHub for the main AMP repo. AMP will host repos. We use that for a lot of other stuff. We'll probably switch soon. But I don't really even go to GitHub anymore. We don't think about GitHub. It's just somewhere that we use as a Git remote. And it's not like there's one big bang where we said, all right, let's all get off of GitHub. It just became less and less relevant in our workflow. And so all of these things, they die with a whimper. And I think we've seen that for a lot of things. Like you can go back to the very start of agents when everyone was all about the tool permissions and even reviewing edit by edit. And people said, well, I'm never going to do that. But you slowly trust the agent. You see that you are the bottleneck and it's not a good use of your time to be reviewing 100 different edits. and finally you come to realize that. So all these big changes, it's not going to be like someone else is foisting it on you.
It's going to be like you kind of realize and then you realize it doesn't matter as much and all these things die with a whimper. And I think that's good because for all of the heat that these things generate when people talk about them, actually it does feel like a pretty real trend that is hitting everyone. And I don't think you see people going back once they've gone to this new way of building with agents. I think that if it was a fad, you'd see a lot of people trying this and then going back. And I just, I don't see that. It's a one-way door. No, it's a total change. It's the way I describe it to people who don't really understand what's happening in tech around AI because I think we're the most highly impacted right now. I got friends that are attorneys and they run their partners at law firms, and they're still, like, they've got time. They're not being forced, like, well, I guess forced, not like negative force, but a force by sheer ability to produce more and better is what I mean by force. Like the nature forces of software developers resisting this change to me is like basically impossible
to resist the new abilities we have as teams, as companies, as individuals. I'm really curious though, given the center of gravity that GitHub has been for our industry, Like, they seem to be dying like a very quick but weird death. And I'm not quite sure if it's a death yet because they're not changing. You got lots of friends inside GitHub, and mad respect for everyone there, obviously. But, like, something is not normal, I guess, to say it. Like, the outages, you can expect it because it's just your force on the platform. And it's just toppling with the amount of the gravity they already had as developers go there. actions has had major issues. You've got merges that have just been bottlenecked. What is your, not so much prediction on GitHub in particular, but more like the way that developers are changing. What's the natural thing going to happen next? Not so much like
GitHub die, but more like what's going to change? What do you see changing as we speak? I mean, you've got your own Git hosting, but it's not, I don't know how robust it is, but you've got some sort of foot in that door. Gwyn? Yeah. You know, GitHub hosting is easy. We don't deserve any pat on the back for doing that. It's easy. What is so special about GitHub is the warm and fuzzy feeling you get from feeling like you're building with everyone else in the world. And the open source community of it, I mean, that is irreplaceable. That has changed the world forever. and private Git hosting, yeah, that's very boring. But what I am worried about, you know, it's easy. It's not going to get anyone excited. Like, yes, it's a Git remote and you can push to it. And you can literally just have, like, a Linux box somewhere that's running Git,
where you SSH to it, and then, you know, your Git can go. It's like, that is so simple to do. But what I'm worried about is with this increasing fragmentation, with agents abstracting away a lot of the building, that's all good. How can we take that progress but keep this idea of the global commons, that everyone is connected through code, that we're all building on each other's work, that we're standing on each other's shoulders? There's something really magical about that. It feels like it breaks down borders. It is this incredible expression of human goodwill in the same way that, you know, Wikipedia is. It had every reason to turn into a cesspool, and yet it's turned into something pretty incredible. Yeah, Wikipedia is pretty special. Yeah, and same with GitHub. this idea of why would I open source, not that GitHub invented open source,
but it made it orders of magnitude bigger and wider reach. This idea that you can build on something that someone else somewhere in the world made and that they open sourced, maybe out of the goodness of their heart in some cases, maybe because they wanted to get more reach themselves. There's a lot of different reasons, but it's this common good. And I'm worried about that going away. If everyone goes to their own little huts and all their Git repositories are disconnected and everyone is building things from scratch rather than reusing, I worry about the quality going down. I worry about specialization going down. I worry about just not feeling connected as a software developer. And there's something about software developers where people talk in the public and talk about what they do way more than most other fields. Yes, other fields will go have conferences, they'll present papers and so on,
but everything we do in software is so much more out in the open. I mean, think of a kid who's 11 years old getting interested in programming, realizing that Linux and Chrome and so much of the software that they use, they can go look at the code there. That's something magical, and I'm worried about that going away. I don't know what the solution is, but I want to figure out how we can keep that warm fuzzy feeling that GitHub has provided And that is the most important thing that they done It changed the world Yeah Yeah I think so I'm still sure what's going to happen because if they can't get their uptime in order, they're going to kill themselves just by lack of trust, right? It's not even like, oh, there's not even something better to go to. There's no GitHub takeover. You know, there's no, maybe you might say Codeberg is there or they're doing something around that, or 4GO or maybe Git T or Giddy, however you say it.
I use Git T locally. So that's, again, back to my issue with Orbs is that my code host is local. I don't use GitHub at all. I set up a VM a year and a half ago. It's the same VM. I've got an amazing backup strategy. It's offside to S3 and everything. Like, nothing's going to happen to my Git host locally, but it's here on my compute, super fast, zero latency, accessible from anywhere I'm on a tail net, anywhere I'm on my tail scale network, and it's private to me. It couldn't be more safe, you know, like in this world. And I'm fine with that, right? I'm fine with, like, ejecting from GitHub in those cases just because it benefited me in the way I wanted to build software. It didn't benefit – it wasn't anti-GitHub. It was just a personal choice that fit. and now every bit of code that I have written is all on my own GitHub. It's not somewhere else. Even pull requests.
I think that's where people are going when they see downtime. There's even NPM registry inside of GitT as well. So I can publish to my own NPM registry. It's like internal stuff. So the walls are breaking down, but they're not breaking down for everyone all at once. But then I see you all with AMP code and I see projects. I see secrets. I see environment variables. I see workspaces. I see multiplayer. And I think these guys have got a plan to take over in some way, shape, or form. Is that not your case? Are you not doing something to take over in some way, shape, or form? I think that for taking over how a team works, yeah, I mean, we've taken over how our own team works. We do push to GitHub, but we don't really think about that. And when we're working on something, we'll go make a thread. We'll go make it multiplayer. then other people on the team can jump in and see exactly how the dev server looks, see everything. People can talk. People can, you know, co-prompt the agent. That's a really incredible feeling. The fact that you can, in AMP, go and spin up a brand new project,
go prototype something, it'll make a new repo for you. You just need to tell it to do all these things. It can even run, and very soon it'll be launching custom domains so that you can actually have your site running or your app running. on AMP, it feels like this new way of working as a team, but the missing piece that we've got some ideas around, but I don't think anyone has figured out, and I don't think people are thinking about this enough, is that global connected feeling. Yes, you can go in your bunker and set up your own Git host, or even AMP. AMP is private to your team. It's not something that's shared. You're right. Yeah. And we can all live that way. But I think that as humans, we've all benefited from getting out of our caves and going and making a town and making a society that's high trust and all of those benefits. And that is better than if each of us was fending for ourselves. And I want to find what that looks like in the agent world.
Right. I think that there's something there around how, take that sales force, vibe coding sales force from scratch, that ridiculous example. That is not what you want, and you wouldn't want that for two reasons. One, it's actually really hard to vibe code sales force from scratch. There's a lot of features you need to build. So many parts. Yeah, that's right. And you don't need to build most of the features. But what if there was some 10% core of the software that you wanted to use? Like, you know, you go to Costco and the Kirkland products are all really good. And they're the white label products. They maybe don't have all the features, but they're really high quality. What if there's that core and you could remix that and you could build your own things on that? Your agent could customize it. And that gets us back to a world where there actually is some global sharing and connectivity going on.
Something in that direction seems very interesting. And that is where you would see agents not just changing how everyone is building software, but you'd have your agent help you customize it for your needs then. And then it's agents helping you run and use software. And that's where, you know, what is softer than software? the jellyware idea. The jellyware. I like that. I like that. I had it right. Jellyware is cool. It's all about a name that you will never forget, and it will stick in your head. That's what we want to find, and that is weird, and that's fun. And, hey, we can do that because we're 20 people, and we get to call the shots. I applaud you guys for, you know, I don't know what went into breaking off in, you know, like you had Cody, which was SourceGraphs. So from an outsider's perspective, this is how I describe it. Kodi is AMP, but Sourcegraph's version of it for your enterprise customers. I imagine you learned a ton about Asian hard. Learned a lot. Yeah, I mean, right? You probably just learned a lot about also how if you have a product in AI that is good at a certain point in time,
unlike in the past where you could say it's good at a certain point in time, now let's go make it go to a bigger market. But no, that is the path to obsolescence faster than in any other software cycle. You've got to always stay at the frontier. Cody, a lot of people loved it. Great product for its time, but there was no way in the definition of that product, which was essentially AI code chat for enterprises, there's no way for that definition of a product to have the impact that we knew that AI could. And that was a huge lesson, huge, you know, I think it got us here. So, you know, mistake, I'm not sure. I think that we're all 1% of the way. I mean, I think even on that note there, there's things you learn by going a direction and getting the feels, right? And that might feel vibe-coded. It might feel like, oh, my gosh, I'm actually over my head. I don't even know what I'm doing. I'm so far in the weeds, but it's a good direction.
We just got to figure it out. Yeah. Yeah, I feel like that's what coding may have been for you all because the result of AMP and the company is clear, right, what you all have been able to do. And I think, you know, as I asked you the question around GitHub and how things will change, I feel like GitHub was a response to workflow. At the time, it was 2009, 2008, the workflow was open source is moving fast. Everyone is desperately trying to keep up, and GitHub was moving at such an inertia because they opened the doors to social coding. It had never been done before. So it was this new workflow thrust onto developer land. And so they gathered up all the inertia of that, and that is what GitHub is still today. The workflow is changing. The way teams think about a new idea, and you know the Amp, think about an idea, the speed to creation, the speed to something real is now that bottleneck. And the way teams are working around agents is 100% changing how we ever written software in the past.
We're still using all the same primitives, languages, packages, et cetera. We're still using machines and compute and RAM and cloud and all those things, but the way we do it now is dramatically changed. And so I think for that reason alone, the workflow forces the entire market to change. And that's what we're doing right now. It's a replatform. You know, as I was telling you before, like when I described to my friends what's happening who are not in tech, the entire state of tech is replatforming on AI completely in every way, shape, or form. Love it or leave it. Graybeard upsetness or not. You know, some are still – I'm still upset about a couple of things, but for the most part, I'm like, you know what? willingly going into this new world, but also just a little trepidation as like, it's all token, it's all cost. The cost to produce software now is, you can still handwrite it, it's going to be a lot slower, but you have to... It's not competitive in the marketplace anymore, so I actually don't think that's an option, just to be blunt. Every, every, you know this because you've taken funding,
every company out there that's taken funding is for sure not handwriting code in this day and age. 100% not. Yeah. It's not the way we do it. Well, we've not taken funding on AMP. So we are profitable. Oh, you haven't taken funding. And congratulations. I didn't know that. Wow. Yeah. So it's very nice to be at a company where the bank account gets higher as you go. And that's part of why we're able to do these crazy things where I love VCs, But they are essentially fuel on a fire. And what we've learned is you've got to, you know, with things changing so fast, the old playbooks just do not work. And, yeah, it's not that we'll never take funding, but we're profitable right now, and that gives us this incredible freedom to explore, to change, to rip things out. We don't have an enterprise sales team. That's another thing. I think that maybe a lot of developers don't realize this, but so much of a company's decisions and product strategy is driven not by the founding team or the developers, but by how can you make a product's features be legible to an enterprise seller and marketing team that's several steps away from the product to go sell it to customers that are looking for a kind of median product.
that informs so much of it and that waters down how you can express yourself through a product and how you can create totally brand new things like you know the idea of orbs if we had a big enterprise sales team and if we were selling to the median enterprise buyer then we'd call them cloud agents because if we call them orbs every single sales conversation that someone who didn't really care about the company and product would have with someone else who didn't really care about the company and product would start with, wait, what are orbs? And the feedback would be, can we just call it cloud agents? And so it's all these little things where we are able to do things differently on AMP and be a little bit weirder. Yeah. Yeah. Well, I had no idea. I should have asked that before if you'd taken funding. I just assumed that you, because of your past in Sourcegraph and you took funding there, I just assumed that you would have just transplanted that same resource set to that. But I was wrong.
That's cool. Yeah. We took a small amount of money to get started. And obviously, we benefited from the trust that Sourcegraph has built up with these amazing customers. And we would not have been able to do the crazy things that we do on AMP without that trust. Like, there's a big bank where, you know, they took a look at what we're doing with AMP, and it looked crazy compared to what they would normally approve, and this was last year. But as we were having those conversations, it was clear that what was going to be able to be approved inside the bank was moving more and more toward what we were saying. And so, you know, essentially they said, these guys seemed right. And there's no way we would have had that trust if we were a brand new company and did not have that relationship. So we benefited a ton. And it's all these unique things about AMP that let us be a little crazier. Yeah, you're definitely the kind of like the rogue set of individuals. And not so much in a bad way, but like you're trying to really, you said before
with fire, my thought when you said that was you're trying to figure out where to put the fire. So I'm not sure how to use that analogy well, but like you can't put VC fuel on a fire that you're like, well, we're trying to figure out how this fire actually works. We're trying to figure out the primitives that made this fire even possible. And I think that's a lot of what you've done with AMP. And I applaud you. Like the way you've done it, there's been a lot of confusing things over the years, over the time of AMP, but it's not like you weren't trying to figure out how to fold all of that into the right user experience. And now, like I've never actually used AMP code in the URL ever until probably the last two months. and that change for me alone helped me see a whole new way to think about as a developer building with agents because building with agents on your local machine is anxiety inducing. You know, I just, that world is not a world I want to live in.
I want to live in a world where my agents have a home, right? I know how much maybe resources and RAM or compute they're sitting on potentially even. Like that to me is like the kind of developer I am. I like bare metal. And so I don't really need – I don't get that necessarily with AMP, but I get it in that I understand your team and how you work. So it's one step removed from the bare metal. But that's kind of where I want to live in is I want to understand where my agents have a home and they have clear workspace they can kind of hang out in. They've got access to Git that's not toppling over or can't merge code or can't do actions because the system is under load. That's the world I want my agents. Almost like I'm farming little kids. You know, like these little agents are like, my little, my little, I got an agent that's three months old. I got one that's five months old. I don't know. I'm just kidding around. But I want to know where they live. I want them to have a home and that home be not on my machine, period. Yeah. More none. Yeah. And there's also something about having all that history going back, you know, from when you started a project,
all that history tracked all the human ingenuity and intent remembered. And, you know, what's the alternative? Look, there's other good products here, but I don't think that if you use Cloud Code, well, you're missing out on so much else. If you're using any other agent from a model company, you're going to be limited there. If you're using an agent that is clearly not trying to track the frontier, but they're probably trying to go get acquired or sell, then all these things, you're setting yourself up for a big change in the future where you're going to lose all of that precious human thought and intent that went into what you're building. So what you get with AMP is it's a team where we care, we use our product all the time, we got skin in the game, and we want to bring everyone along and explore. And with AMP, we want it to be impossible to use it in the wrong way.
We want it to be impossible to find some feature where we're like, oh, actually, no, that's not good, but we had to keep it in for this customer. You're always on the golden path with AMP. I love that. And we are going to be around. We've been doing this. And every single day we wake up and it's just, you know, we get to explore a little bit of the future with all of you. The frontier. To the frontier, my friend. Maybe one last check in if you have a minute. I think you probably got one more minute left for me. We should check in on the tail net thing. Let's see what happened there. Let's close that loop at least for the audience and even me to know if I'm a happy AMP user or a really happy AMP user. One of the two. Usually I would be checking in on this, but I gave this my full attention. And let's see. Here's my prompt. And I'm back in AMP looking at when I asked AMP to go make it so we can solve Adam's problem so that AMP orbs could work and clone his repos that are only on his tail net. Yes. I'm reading this.
I can see I'm solving the problem. I could probably – the issue is that you instantiate the orb assuming that a repository doesn't need a credential. Yeah. That's the problem. So you need that credential sooner than your orb instantiating the Git repo. So it's a sequencing issue. Yep. So let's see. It also updated the docs. So we've now installed Tailskill and AMP orbs. It seems like we've got enough customers that want this. Actually, I think we're doing that already. Yeah. All right. So, yeah, the key thing is now we start tail scale and we join the tail net if the environment variables are present before it clones the project repository. So, hey, man, that's what we need. I'm going to put my secret in where you put your secrets. So you already have a secrets vault. I want to put that tail scale API key in that secret vault that you already have for me. And when I spin up orbs, I want you to use that key before you instantiate your orb and your Git repo and you pull from the tail net. All right. Let me just kick it off on a verification. I forget if I have Tailscale set up myself, but we'll see.
I'm going to dictate to AMP now. I'm going to say, check to see if I have any Tailscale credentials set up already, and if so, then go and make another machine or something somewhere that exposes a repo that can be cloned only via the TailNet, and then go try what you have in your local dev server and try cloning from that. There we go. So it's going to go and test and verify. we'll have to close the loop maybe another way. I know you've got a board meeting to go to. I appreciate you going longer and extra for me. I know you did. Audience, you don't know that, but I know we're over on time. And Quinn, massive fan, mad respect. Love what you guys are doing. Don't stop. Keep going. And that's it. Let me say goodbye so you can go and do your board meeting. Yeah, it's Orban time. It's time to get back. Yeah, thank you. Thank you. What stays with me from this episode is Quinn's last line, that every thought a human has had about what you're building is precious.
Agents keep getting faster, they're getting cheaper, and the model matters less than it used to. What's scarce now is intent, what you actually wanted written down somewhere and agent-defined somewhere among them. That's why threads matter and why review is shifting from reading every line of code to checking what the work actually does. There's a tension in there, too. Quinn worries that as we all move our code into private corners, our little worlds, we lose the shared commons that open source built. And dare I say it, the corpus these models have been trained on. A few questions to take with you. When your agents have a home, where does your intent live? Who else gets to build on that intent? And if more of our code goes private, what will the next generation of models and humans learn from? You know, I've got some awesome sponsors that take care of me, that make this show possible, one of which is Parallel. Parallel.ai, that's where your agents find answers, search API, extract API, decisions API, Parallel, MCP.
These are all web tools I've been using. From Parallel. And our friends at WorkOS, every login, every auth I deliver uses WorkOS. AuthKit is amazing. It's so cool. And it's free. Up to a million users. 100% free. And, of course, to my friends at DeetGram. My gosh, my transcripts are now more amazing. My intelligence for this podcast is more amazing because DeetGram is amazing. Voice AI for real conversations. Learn more at dgram.com And of course, the Beat Freak in residence Brake Master Cylinder BMC, bringing those beats Nodding our head, loving it Alright friends, that's it This show is done We'll see you next time
Winning. It really whips, though.
番組の概要欄(原文)
Quinn Slack is back and we're talking about the next level for Amp. Agents write the code, they review the code, and they can push it to main (like our friends at Amp). But where should your agents live? Who remembers what you actually asked them to do? Amp is now its own company, and Quinn has a lot to say about orbs, threads that last forever, and why code review as we know it is fading. He explains why the model matters less than it used to, and why he worries we could lose the open source commons. Quinn even built a feature request live on the show to solve my tailnet problem.
出典・この記事について
- 番組
- Changelog Master Feed(Changelog Media)
- 元の回
- Orbin' with Amp (Changelog Interviews #691)
- 配信日
- 要約作成日
- 公式サイト
- changelog.com/master
この要約・解説記事は、元の音声を文字起こしした内容をもとに AI(Claude)が日本語で作成したものです。 固有名詞や数値に誤りを含む場合があるため、正確な内容は元の音声でご確認ください。制作方法について
関連エピソード

MCP、MHS、そしてオートメーションピラミッドの終わり?
Industrial AI Podcast

オープンソースはAIエージェントを安全にできるか?
NVIDIA AI Podcast

トークンを買うか、AIファクトリーを築くか:エージェント時代のトークノミクスとオンプレ推論(Sadeepa Wijesekara氏)
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)

企業のAI活用はどう変わっているか:トークン使用量は36%増、支出は2%増
The AI Daily Brief: Artificial Intelligence News and Analysis

技術面接は壊れている:HackerRank CEO が語るAI時代の採用
Software Engineering Daily

ジェンスン・フアンがNvidia CEOを辞任、AnthropicのIPO、AI 2027の前倒し予測、HubSpotのAI再創業
The Artificial Intelligence Show