
マルチクラウドLLMルーティングと、コンテキストを理解するコーディングエージェント
Multi-Cloud LLM Routing and Context-Aware Coding Agents

マルチクラウドLLMルーティングと、コンテキストを理解するコーディングエージェント
Agentic Conversations (formally mlops.community)
要約
CloudflareのMatt CareyとMiroのScott Crooksが、自社のAIゲートウェイとエージェント基盤、AIコードレビューの運用を語った回。Miroはリージョン、クラウド、データ居住要件を束ねてモデル呼び出しを振り分ける。Cloudflareはエージェント専用のコーデックスを使ったレビュアーをFlue上に構築した。オープンソースの運営負荷、社内ツールの公開と終了のさせ方、コードを書く量が減るなかで生まれる「学究的エンジニア」像にも話が及んだ。
この回のポイント
- ●MiroのAIゲートウェイはモデル呼び出しとビジネスロジックを担う。たとえばドイツの顧客にはEUのモデルだけを使うといった制約を組み込み、リージョン・クラウド・直接提供元をまたいでリクエスト数やトークン数の上限を集約してルーティングする。
- ●Cloudflareはエージェント向けに、AIゲートウェイ、Agents SDK、Astroの開発陣によるJavaScript製フレームワークFlue、オープンソースのエージェント版オフィススイートCloudflare OSを提供している。
- ●CloudflareのコードレビューはFlue上で動き、エージェントだけが読む社内エンジニアリング・コーデックスで会社の流儀を共有している。Matt Careyは、これまで使った中で最良のAIコードレビュアーだと述べた。
- ●MiroはEaselbot、通常のClaude、自社のAIレビューボットの3層でレビューし、さらにOplaneでセキュリティやコンプライアンスを確認している。Easelbotは本番ログから夜間に修正PRを作るが、マージされるのは約半数で、その結果を学習に回している。
- ●AIによる安価なPRの急増に対し、MCP SDKは外部の非共同作業者によるPRを止めた。Astroはissueをゼロにし、これから常にゼロを保つ方針だという。
- ●コード生成が楽になるなかで、コードを読む量より概念や産業への理解が重要になるという見方から「学究的エンジニア」が語られた。
章立て
AIゲートウェイとは何か
Miroのゲートウェイは、モデル呼び出しと、EU限定モデルのような顧客要件を扱うビジネスロジックを一つにまとめたものだと説明された。
マルチリージョン・マルチクラウド
同じモデルをAWS、Azure、GCP、Anthropic直接など複数経路で提供でき、上限を集約して経路を決める仕組みを紹介。Sidekicksでは自動モードも提供する。
CloudflareのエージェントスタックとFlue
Agents SDK、Flue、Cloudflare OSの概要と、JavaScriptが動く環境ならどこでも動かせる点が語られた。
コーデックスを使ったコードレビュー
エージェント専用の社内コーデックスと、Flue上のレビュアーを紹介。エージェント体験チームとベンチマークの話にも広がる。
Miroの3層レビューとEasel
夜間にログから修正PRを作るEaselbot、Claude、独自レビューボット、Oplaneの役割。PRの約半数がマージされ、その結果をフィードバックとして使う。
ソフトウェアファクトリーとOSS
Astroのissueゼロ化や、低品質PRの選別、外部PRの停止など、OSSメンテナの負担軽減策が語られた。
社内ツールの公開と終了
社内ツールを公開する際の評判や維持の責任、静的な時点として明示することや寄贈、安定化による終了のさせ方を議論した。
ホットテイクと学究的エンジニア
チョコとナッツの雑談を挟み、物理世界とソフトの融合、コードを読む量が減るなかで概念理解が重要になるという展望で締めた。
解説記事
CloudflareのMatt CareyとMiroのScott Crooksが、それぞれの社内でAIエージェントをどう使っているかを語った回である。話題はAIゲートウェイ、コードレビューの自動化、オープンソース運営、そしてコードをほとんど書かなくなったエンジニアの役割まで広がった。
モデル呼び出しは単純ではなくなった
Scottによると、MiroのAIゲートウェイには二つの役割がある。一つは実際にモデルを呼ぶ機能で、自社ホストのモデルと通常のLLM呼び出しを併用している。もう一つは顧客対応のためのビジネスロジックで、たとえばドイツの顧客には欧州のモデルだけを使う、といった条件を組み込む。
複雑さの源はモデルの提供形態の多様化にある。当初のAWSでは特定リージョンにしかモデルがなかった。いまはリージョンに加えてクラウドも複数になり、同じモデルをAWS、Azure、GCP、あるいはAnthropic直接で提供できる。Miroは経路ごとに毎分のリクエスト数やトークン数の上限を集約し、その枠をもとに振り分けている。AI機能「Sidekicks」(Miroにおけるエージェントの呼び名)では既定モデルに加え、コストを抑えたい顧客向けの自動モードもあるという。
Cloudflareのエージェント基盤とFlue
MattはCloudflareの構成を紹介した。外部モデルやWorkers AIのホスト型モデルへ振り分けるAIゲートウェイ、自作向けのAgents SDK、そしてAstroの開発陣が出したFlueである。Flueはエージェントの制御フローを書くためのJavaScriptフレームワークで、Node.jsやCloudflare Workers、ブラウザなど、JavaScriptが動く場所ならどこでも展開できるという。ツールの遅延読み込みといったハーネス設計もしやすいと述べた。
さらに、オープンソースで自分でデプロイできる「Cloudflare OS」も話題になった。エージェント版のオフィススイートで、スライド作成や、部署ごとのダッシュボードをブループリントとして共有する使い方ができるとされる。Matt自身は家庭の食料品の買い物に使っているという。
コードレビューの要は会社のコンテキスト
会話の中心はコードレビューだった。MattはFlue上に作った社内のレビュアーを、これまで使った中で最良だと評価している。外部公開の予定は未定だという。成否を分けるのは、会社として何を良いPRとみなすかという集合的な文脈の量だという。
そのためCloudflareは約1年前から、エージェントだけが読む社内エンジニアリング・コーデックスを整備している。使ってよいツール、例外の扱い、課題へのアプローチなどを、会社の運営に関する調査をもとにエージェント自身が書いた。かつてAirbnbのスタイルガイドがJavaScriptの書き方を方向づけたように、このコーデックスでアプリ全体を作れるようにしたい、というのが彼の狙いである。
関連してDemetriosは、ドキュメントをエージェントが理解できるかを評価する難しさを挙げた。Mattは「エージェント体験」が新しい分野になっていると述べ、Cloudflareの専任チームが、最先端のコーディングエージェントがCloudflareのAPIをどれだけ使えるかのベンチマークを作ったことを紹介した。
Miroの3層レビューと、半分しかマージされない意味
Scottによると、Miroのレビューは三つの仕組みで構成される。Easelbotは会社全体の文脈とアプリのログにアクセスし、朝には頻発エラーに対する修正PRが2〜3件できている。個々の変更には通常のClaudeによるレビューがあり、パイプラインで自動実行される自社のAIレビューボットは、Easelbotの修正案に対する人の承認や却下を見て、人が重視する点を学ぶ。
興味深いのは、開かれたPRのうちマージされる割合がおおむね5割だという点だ。Scottはこれを重要な指標とし、関連性の判断を強化学習的に返している。セキュリティ面ではOplaneを併用し、GitHub Actionsの設定やシークレットの露出、データの伏せ方などを確認しているという。
OSSの負担、社内ツールの公開、学究的エンジニア
Mattは、Astroがissueをゼロにしたと紹介した。ゼロを保つ方針だという。一方、AIで安価に作れるPRの流入は、オープンソース運営の負担になっている。Superpowersの作者は低品質なPRを自動で選別していると語り、Mattが関わるMCP SDKは外部の非共同作業者によるPRを止めてissueのみにしたという。メンテナにはプロジェクトの方向性への強い考えがあり、従来のやり取りの手間をどう減らすかが課題だと語られた。
社内ツールの公開については、Kelsey Hightowerの講演を引いて、評判の面からも公開した以上は維持が求められるという議論があった。LLMでデモから同等物を作れる時代には、公開コードを「ある時点のもの」として明示する、他者に寄贈する、安定版にして開発を止める、といった終わらせ方が重要になるという見方も出た。
終盤でMattは、コードを書く量も読む量も減るなか、システムや産業の概念への理解を深めることが重要になると述べ、「学究的エンジニア」について記事を書きたいと語った。Scottの同僚はこれを「紳士科学者の時代」と呼んでいるという。抽象度が上がるほど、適切な言葉と概念を持つことが効いてくるという考えである。
まとめ
日本の読者にとっての注目点は三つある。第一に、モデル提供の経路が増えるほど、レート制限やデータ居住要件を束ねるゲートウェイ層の価値が高まる。第二に、AIレビューは導入するだけでなく、社内の規範を書いたエージェント向け文書と、マージ率などの人の反応を学習に戻す仕組みで精度が変わる。第三に、コードを読む量が減る時代に、エンジニアには概念理解と取捨選択が求められる。自社の運用を設計する際の具体的な参考になる回だ。
文字起こし(英語・自動生成)
The biggest use case internally is we have this, I think, the best AI code review I've ever used. I have no idea when we're ever going to share it externally or if we ever will, but it is amazing and it's built on flu. Is it primarily for humans to read? It's only for agents to read. When people approve or reject, our AI review kind of looks at that and says, okay, this is not only what I think needs to be fixed, but what is important to people. Because that's the more interesting statistic that we found with our internal tool is how many PRs get opened, how many get merged, and how many get closed. PRs are cheap, man. PRs are very cheap. Yeah, it depends on your model, but yeah, potentially they're cheap. And maybe this is super sad. Like, I think agents are bringing us into the terminal and away from people. You've got to get out of here people and do things like this, yeah? I think my first hot take is... Almonds are the only nut that you should eat with chocolate. We're back.
I'm with Matt and Scott. How are you guys doing? Yeah, pretty awesome, eh? Awesome. Dude, I want to talk about so much with you. You're at Cloudflare. You're at Miro. There's so many different things that we can go into. I think we were just talking about AI Gateway, right? Yes. And it's funny, the last conversation I had with the Booking.com team, they were talking about the MCP Gateway that they put together. What do you define as an AI Gateway? Yeah. I can tell you how we define it. I'll give the maybe like the quick understandable version, which is similar to I think a lot of companies did this. It's, hey, we need something really quick. Let's just throw it together. Yeah. So it was whatever use cases were amalgamated to create that. For Miro, what that means is it's two things, really. It's an AI gateway that actually does the model calling, right? So, like, there's the back end, let's say, of models. In our case, we have a good mix of self-hosted versus just regular LLM calls.
And then the other part of that is the business logic. So Miro has some very specific things that we need to do in order to make sure that customers are happy with our product. A good example of this is you're in Germany. You say, I only want to call European-based models, right? That's a very simple use case. We code that in. And so these sort of things is what's kind of self-contained. Is it also routing? It's also routing, yeah. So one of the things that we've seen is an evolution in how the models, particularly with the cloud providers, has gone. We started with AWS. You would have models that were really only in certain regions, right? And so as they were also figuring out their infrastructure, we said, okay, let's route everything to this region, get everything working. As we've gone along, we've seen sort of dimensions come along. So we've not only seen multi-regionality, because we deploy in multiple regions, but also multi-cloud. So this is like we have Opus 5. We want to deploy this not only in all the regions that we can within AWS, but guess what?
You can also serve that through Azure, through GCP, or Anthropic Direct in certain cases. And so it becomes this sort of like three-dimensional problem of like multi-region, multi-model. in the end what our model gateway does is you know we have a pool of providers that allow us to essentially define okay for this model what kind of requests per minute can we get what kind of tokens per minute can we get and all of that just gets aggregated and then we make routing decisions based on this dude i totally forgot back in the day where you could only get open ai through azure exactly and you saying that reminded me of like oh yeah and when you remember when you would get rate limited yeah like after two prompts you're like what happened all the time and so now yeah okay you've got these as you call it multi-dimensional and you have not only for the same service if you're dealing with the anthropic models but then i imagine you've got these other models that you're dealing with yes and so maybe you've got some open source ones and are you routing depending on the complexity of the tasks uh we are yeah so miro has a program
or a capability called sidekicks. It's a fancy way of saying AI agents. And it does exactly this. So we have default models defined, and then there's sort of a layer that also does an analysis and says, okay, the customer requested this, but they also, let's say, want to maintain costs, and they talk to us about this. So we also give them some logic around, yeah, like kind of an auto mode. So, yeah, this was also something that came about pretty quickly, and because we have the model capabilities and the pool, as I said, it allows us to do a little bit more complex things on the back end. You guys built something like this too? Yeah, we have an AI gateway for customers. Yeah. And you came out with the OS recently too, right? Can you tell me more about that? We have like a full sort of like agent stack. So the AI gateway can route to external model providers and come out to like the workers AI hosted models. And then we also have like the application layer on top.
So we have the agents SDK, which is for people building their own stuff. We have Flu, which is like a big old framework for building agents. It's kind of taking a new devX around. It's like from the, from the makers of Astro. It's really, really interesting actually. And then we also have this brand new thing. I was going to go through all of them, but this brand new thing, which is like this full application that's open source that you can deploy yourself. and it's like the uh uh you can think of it's like the off the a uh agent office suite for your company and we call that cloudflare os so i have a version of it deployed for like um literally like me and my my girlfriend as well and we like we have it in in our house uh which is kind of fun uh but then you can also do what it's in the house oh it's like just like shopping stuff like like like literally food shopping it's like the easiest one but um people like come like there are a bunch of companies i have no idea which ones i'm going to name so that have it deployed like internally as their like all-in-one office suite you can make slides out of it you can build these things these apps like blueprints for these apps so you can have like your finance team building like
special dashboards and then you can save them as blueprints and then you can other teams can spin them up you can have like uh skills by by like job title and things like this and like job and It's really, really fun, and it's built on this capability model where the agent only gets access to the app that it's looking at and, like, all internal to the Cloudflare network. And what's the Astro one that you made? At Flu, the Flu framework, yeah. So we released that. Well, the makers of Astro sort of went out on a tangent, actually, and released that, and it's really, really cool. It's a dev experience. It's kind of like a React, like, functional dev experience to build agent control flow. and it just natively deploys to anywhere that JavaScript runs. So you can deploy it to a Node.js server, you can deploy it to like for self compute or you can deploy it to Cloudflare workers Ooh okay you played Yeah Flu Yeah yeah yeah The Flu I work on Flu like a bunch So yeah we running a bunch of production workloads internally
We hack on it. We build stuff with it. We build applications on it, and we give applications to our customers on it. What have you been building? Like, I think, like, the biggest use case internally is we have this, I think, the best AI code reviewer I've ever used internally to Cloudflare. I have no idea when we're ever going to share it externally or if we ever will, but it is amazing and it's built on Flu, yeah. Why built on Flu with the Code Reviewer? Is it just like reading these files? It's reading your commits? Yeah, all of the above, and it's got its own custom tools. You're building your own harness. Some people might want to run Cloud Code in a container, or you might want to run like open code on a durable object or like Pi. You can run all of these things, but it's very much like a lower level thing. Or you can, Flu is a whole framework. So you write in JavaScript, but you write like Flu semantics and then it'll deploy wherever you want to deploy JavaScript.
And you can specify all of these details. Yeah, like you can specify tools, prompts. It's really the innovative thing with Flu is like the control flow. so it's very easy to do things like lazy loading of tools and like more interesting harness engineering stuff because of the way that uh fred and the astro team have done the control flow yeah i see it's it is really good fun it's really fun and it'll run anywhere that javascript runs so it's not just like tied to cloudflare you can you can run it wherever yeah even runs in the browser oh nice dude well as you're saying that it just makes me think like man i haven't gone deep on personalizing my coding specifically like the PR reviewing agents because I get so much lift with the ones that come out of the box how much of a difference do you feel like there is between those two when you're going in and you're tweaking everyone versus the just like yeah the
skill that you get off the shelf maybe it's a ponytail skill or something yeah i think it's all about like how much collective context you have about your company um and about like what your company denotes a good pr i would say um at least that's the way we started uh classless cool in a few months ago well actually almost a year ago we started developing this uh internal engineering codex which i think is public now but it's like how much it's like what denotes um what what what tools we're allowed to use, what tools we should be using, how to make exemptions to this, how to think about approaching an engineering problem at Cloudflare. And a lot of other companies have similar versions to this. I know the very famous ones are the Airbnb style guide for code, right? People used to write JavaScript in the Airbnb style guide. Now I think you can build whole applications with the Cloudflare codex. I think that's how we're trying to think about it. Is it primarily for humans to read or for... It's only for agents to read. Okay. Yeah, it's only for agents to read. So it's built by agents based on extensive research about how the company operates.
And we're trying to write down how the company operates and trying to be aligned across the whole company. Cloudflow is 4,500 people, I think, and trying to be aligned about everyone's coding agents should know how the company works and how the company operates, where to go for different things. Yeah, it's a massive task. but when if you can do that and you can generate this like codex uh codex is such a loaded word now uh but if you can generate this like this like uh um context pool then you can have really good success with internal agents because the internal agents can also do um can do like best practice yeah one thing that i find super difficult is just knowing how well my stuff's performing versus like what it should be at or what is the top end of this performance and so like evaling it and going through and even just something as simple as documentation and the agent's ability to understand and read the documentation and evaling that you know is something that you
wouldn't necessarily think about when you're making it for humans yeah because it's a human all right like i want to write the best i can i want to make sure that these docs are well written but you're not like having to go and then like say can we eval these docs right it's this whole new field of like agent experience yeah um and i think we had liad uh uh presenting here in mcp agent con um and their whole company is around like aura is around like agent experience there's a bunch of other startups are trying to crack this like anything so it's at cloudflare we have a team called agent experience yeah and they do awesome stuff they maintain all of the they maintain a dashboard agent, but they also maintain, like, every other interface. Well, they help with other interfaces, so they help with the MCP server to interact with things, and we're constantly experimenting on that front. And, yeah, we have evals. The coolest thing I think they did recently was a full benchmark for how well do the frontier coding agents, how well can they use Cloudflare natively, how well can they use the APIs.
Like, do they know our open API specs? Can they deploy a worker from scratch given, like, literally nothing, just intimate access? I think that's really fun. Have you guys experimented with anything like this? Yeah, we have an internal bot that does exactly this. So it seems like there's a common pattern emerging. So we call it – so Miro is the canvas, so we call it easel bot, right? It's pretty cool, actually. Yeah, so what we noticed was with the code review problem, we actually have, like, kind of three checks, and they all are similar, but they all seem to hit different areas, and I think that's why we've kept all three of them. So we have Easelbot, like that is more of like a getting the whole company context, fixing errors where it finds them, but it's also our application that gets access directly to our application logs. So when you wake up in the morning, typically on our bigger, more active code repositories, we'll see, yeah, two or three PRs. Hey, I saw this has been appearing like 10 times in the last hour. This is what I think is the potential fix. So we have that level. And then for the individual reviews, we have just normal Claude, right?
It's like, hey, Claude, can you review this, please? And then we have our own AI review bot. This one is also automated in our pipelines. It's kind of a combination. So I think it takes a lot of the potential fixes that have come from Easelbot. So when people approve or reject, our AI review kind of looks at that and says, okay, this is not only like what I think needs to be fixed, but what is important. to people because that's that's the more interesting statistic that we found with our internal tool is how many prs get opened how many get merged and how many get closed yeah um and so we find like there's about a 50 50 split so right now we're still trying to trying to find that sweet spot because yeah a lot of prs can be open these days i'll tell you what yeah so we and the good news is that like we have still like engineers who are looking into this and And you know we have on people saying yeah this is relevant this is not So it not So it kind of like we experimenting through reinforcement learning Like this PR is not relevant close it and then that just feeds back into this
Oh, nice. So it's nice because we have not only multiple streams for when you're just opening something and then you get the review, but then we also have on the back end like looking at the contents and see if it was relevant or not. So we get it from both sides. In addition, we also use Oplane. That one's a really good security, if you guys have discussed it. Has that come before? I don't know. I don't know what this is. Oplane, O-P-L-A-N-E. It's kind of like a security check, let's say. So you open a PR. It gives you kind of like a breakdown of compliance, and it can really range from anything. So it could be something as simple as, like, we use GitHub, right? So we have GitHub Actions. It could say, hey, you haven't shoppened this action. And so it'll give you an explanation of, like, why this is important. And then after the fact, it'll also get more into detail about, like, your specific change. So it's, like, kind of like an all-encompassing security review so that, like, when you merge it, you feel confident.
Like, okay, like, I've hit or the agent has hit, like, kind of a base. I'm not exposing any secrets. I'm not exposing, yeah, exactly. So it'll give you, like, recommendations on, like, how do you redact data. It really is kind of an all-encompassing tool, so it's great. I think that the upshot of, I guess, this conversation, I didn't realize how much you had internally. That's so awesome. But everyone is building these software automations, I think still, with the intention of them becoming full factories. Everyone's trying to close the loop at some point to become full factories. And I think one of the fun examples that people can go and have a look at because it's very public is how some of the open source repos are actually doing this to close the loop. Astro, the people I talked about earlier making this JavaScript web framework, they got to zero open issues today. Wow. For the first time ever. From literally thousands of open issues, they got to zero. And they are going for inbox zero from now on, which is insane. Because up until now, I think open source repos have really had this deluge of people,
I want to make a contribution, I'm going to throw my agent at this potential problem and see what it comes up with. I'm going to scan the repo and find these, like, security vulnerabilities, which potentially – Well, I know the guy who created Superpowers, that pack, Jesse, he told me that for his open source projects and Superpowers, he was like, I get a lot of PRs for Superpowers. And a lot of times it's very low quality because obviously, yeah. but he built out ways to just like automatically sift through the low quality ones and then the ones that should get the human's time get bubbled to the top by having like some very firm rules yeah i think everyone everyone is doing this like in the extreme cases for a lot of the mcpsdk's i can't speak for all of them but we've actually turned off external uh non-collaborated prs and people just make issues and they can if they want to have a go at making a patch they can link the
patch in their issue and then we can go in and do it i think there's this like whole there's this whole thing about how can you reduce that maintain a burden and how can how can you actually automate each of the points but have some type of human review and like whatever people are comfortable with because prs are cheap man well yeah it depends on your model but yeah potentially that that they're cheap uh but also a lot of the time i think open source maintainers have a very like strong idea of where they want the project to go and so them knowing about issues is very useful but then actually the implementation has to fit with the direction that the project is going and that's often that previously involved like a lot of backwards and forwards uh between maintainers and the person wanting to fix the issue but now it is just cheaper the backwards and forwards is sad and maybe this is super sad like i think agents are bringing us into the terminal and away from people um you don't gotta get out there and yeah you gotta get out and meet people and do things like this yeah we've spent like the last two days in mcp maintainer meetings and it's just been so fun just to like riff on ideas and things when normally it's uh it's just it's it's very much
agent to agent in the world right now you're talking to your llm too much and makes you delusional sometimes you gotta touch grass we'll say i hope that this also has another an opposite effect like a few years back at kubcon it was the one that was here in amsterdam and uh kelsey Hightower who of Kubernetes fame. I went to a talk that he did and it was talking about open source contributions for companies. Right. So it was specifically like, okay, you're coming to this talk because you you're at a company, you want to open source something like, let me give you some advice on that. Um, at the end of the talk, what he said was like, you know, you have to think not only about the, the monetary implications, right. Does this make sense for your company to, uh, put time into open source, but it's also a reputation thing. And so he said, you know, I used to work. He mentioned a company. I won't mention it here. He said, I used to work and consult for this company. You know, they were very famous very early on when the microservice kind of trend was taking off. You know, they released a lot of products, a lot of libraries, things that took off, and
they didn't maintain them. I hope in the future that this will also kind of encourage companies to take something that they really are developing internally and say, you know, this is something that should be public. The more recent one, the talk from Spotify just a couple hours ago, this is exactly, they're like, hey, we have this MCP gateway that we want to internally have people use. The fact that they're developing this, they already had success with Backstage, and so I think people are now like, please, yes, like, release it. So having this agentic kind of world also, I hope, enables that to get more quality products out the door. And there's certain companies that it feels like it's in their DNA a little bit. I know that before backstage even, Spotify put out, was it Luigi in the data world? I don't know if you're familiar with that. It was like an Airflow competitor. And I think the guy that wrote it, do you all know Modal, Eric Bernharsen? Yeah, yeah, yeah. I think he created – I want to say the name is Luigi.
Somebody can correct me in the comments. But it was that. I've heard this story. I've asked him about this as well. He was working at Spotify, and then he put that out, and then Airflow went out a little bit before or a little after and then kind of took all the air out of the room. What I'm trying to say, though, is you've got some companies that are very much like, yeah, cool, let's release it, let's put it out there, and let's make it a thing. And then other companies, I think they're learning how to do open source because maybe it's not as big of a priority. And I guess I would say, like, the LLMs are making – allowing people to generate vast amounts of code very, very quickly. Whether your company wants to support that code in the long term is interesting. And maybe the value of like fun ideas goes down or like fun demos goes down and like previously you see a fun demo and maybe they didn release the code and you were like maybe a little bit bummed out by that because you like oh I like want to see that but now like that so true I can just make myself now
like from the front demo you can make it yourself so maybe they don't need to release the code maybe it's still helpful but if they do release the code then it's worth like marking it as like this this is like static in time like this is a yeah this is a this is a moment and like we're not going to carry on from that and i think a lot of companies who have like an open source direction like uh cloudfly we have like a bunch who work on open source i work on the mcpstk like a lot of time um we it's very hard to work out like what what you are going to support a lot longer term and also how you deprecate things in a nice way for customers like how can you be like this was a time and a place but actually the whole industry has moved on right now and this is going to be the last thing, guys. Yeah, how do you set that expectation? Can we donate it to someone if people are still using it or can we have, like, we don't want to support this anymore? I'm not going to name specific projects, but, like, that is definitely a thought process. It's a life cycle of a lot of different projects. Or can we make something stable so that people who are using it, they can carry on using it, but, like, we're not going to continuously develop this.
Yeah. Well, speaking of which, that reminds me of the idea that I wanted to get after with you guys, your hot takes. You got any good ones for us? I can give you mine. Go for it. Wait, I have a whole list of them on my phone around here that I wanted to share at some point. I think my first hot take is almonds are the only nut that you should eat with chocolate. Walnuts, no good. That's so hot because hazelnuts. Yeah, hazelnuts, no. It's too much chocolate on chocolate. I don't like it. I guess does, like, the whole Dubai chocolate craze, like, fall out? Fall into that? Exactly. I'm not a fan of the Dubai chocolate. Again, almonds are the nuts. That's my hot take. Yeah, that's too hot for me. Hazel nuts work so well. Yeah. No, I'm out. I do have a hot take, and I have to give credit to some of my fellow mirror engineers. There's a lot of us here. So it was, let me see, it was Sanket, it was Svetlana.
Shout out. Yeah, and Murat. So we were – hot take, it's more of like we're coming to this conference and all of us are from different backgrounds, different countries. And I think the one interesting thing that we all sort of came to the conclusion of is that because now software is so easy to write, the time – you're a little bit free to think of other things. And one of the things that we all agreed on was the combining of something in the physical world with software now becomes just that much easier and will make life that much better. Not only that, but being somebody who is maybe from a different background in software. So, for example, my background is chemical engineering. Oh, nice. Oh, mechanical. Yeah. So it's the same. You come from this background. It's like, well, what does that have to do with software? And it's like, well, some would say nothing. but but others it's like yeah but you're also learning kind of like generic concepts that you can then apply to to other domains right and and that's the power of it yeah we mean to write this
article about like the scholarly engineer which is like which is like i guess as like i'm still a software engineer even though i write less code than i did previously i also actually i went through a phase of being like oh but i just read all the code um and i that's like how i like imprint my mark on it um when the lms are writing and now i'm not even that pro that idea like i read actually much less of the code but i still understand the ideas and then there's going to be a point where maybe like i understand less of the ideas and so that's the bit that i'm like struggling with where if i'm not reading if i'm not uh writing the code um and i'm reading less of the code at least um i'm doing less code review than i was before uh and i'm leaning much more on like my own understanding of how the system works then actually in those previous two times when i would have been like like hunched over a laptop like writing and reading uh i should now actually just be learning more about the concepts of my industry the things that i can build upon like knowing how to prompt is actually just like knowing more about the world that we live in and like
knowing more about like given this situation i should be thinking about this as an idea like that the scholarly engineer is like something i'm really like it's a higher level concepts that you need it's almost like the abstraction is going up and so you're going to be much more effective if you understand what you need to use the right tools you need to use or the right words you need to say in the right moments similar my team calls it uh the his he calls it his gentleman gentlemanly scientist era which i find hysterical and it's like this vision of this uh like gentleman scientist with a big top hat and maybe a pipe just pontificating about the world and like reading like massive books and yeah philosophizing about philosophy yeah they're yeah they're like uh their their impact and then doing very very small experiments that actually have like exponentially large results yeah exactly that's the key is like all right well gathering
a bunch of inspiration from random places around the world and around like the the educational sphere and then implementing that into what you're working on because uh i can't remember where i was reading it but it's pretty obvious that like being able to take ideas from outside of your little bubble and then implement them in your bubble is going to give you much richer results yeah perhaps it'll be in the future where you know instead of having like one career path and saying oh i'm going to be this uh for my for my career it'll be more like what what problem do i want to solve or problems yeah and then it'll be looking at the space and saying what information do i need in order to get to solve that problem yeah kind of combining things at the very end you heard of the idea of like general managers uh general managers are a problem i'll send you afterwards but it's uh it's more of a philanthropic viewpoint of like there is there are these large problems in the world like i want to be the general manager of this like singular thing yeah
like yeah i don't know um i i wish the lms and i hope the lms move us more towards spending our time upon these like more philanthropic is maybe the wrong word here but like that's all yeah good for the world like ideas yeah um you get to level up and that's the cool part is that you're not spending your time being there which some people love i will say that like so if you love it it's just like you know working on old cars whatever or your motorcycle it's not totally necessary but if you love it do it go for it otherwise yeah you can abstract away and uh hopefully get these bigger ideas and and find be able to chase what you enjoy so i think with that man that's a great plate to end it. Feels like a good note. Nice. I feel warm and fuzzy inside. Yeah, it was good. Guys, awesome for doing this. Thank you so much.
番組の概要欄(原文)
<p>In this episode, we sit down with <em>Matt Carey</em> and <em>Scott Crooks</em>, engineers at <a href="https://www.cloudflare.com/" target="_blank" rel="noopener noreferer">Cloudflare</a> and <a href="https://miro.com/" target="_blank" rel="noopener noreferer">Miro</a>, to look at how two very different companies are building their own AI software factories, and what happens to engineering when the code mostly writes and reviews itself.</p><p><br></p><p>We start with the AI gateway: why a model call is no longer simple when the same model is served across multiple regions, multiple clouds, and direct from the provider, and how Miro pools rate limits and data residency rules (like EU-only models for German customers) into a single routing layer that also powers Miro’s Sidekicks. Matt walks through Cloudflare’s agent stack, from the AI gateway and Agents SDK to Flue, a JavaScript agent framework from the makers of Astro, and Cloudflare OS, an open source agent office suite he even runs at home for grocery shopping.</p><p><br></p><p>The heart of the conversation is code review. Matt describes what he calls the best AI code reviewer he’s ever used, built on Flue and powered by an internal engineering codex written only for agents. Scott breaks down Miro’s three-layer setup, including Easel, a bot that reads production logs overnight and opens fix PRs before anyone wakes up, and explains why only about half of those PRs get merged, and how that signal is fed back to teach the system what people care about.</p><p><br></p><p>Along the way, we cover agent experience as a new discipline, how open source maintainers are handling a flood of cheap AI pull requests, Astro reaching zero open issues, when companies should open source internal tools and how to retire them gracefully, plus a few hot takes on chocolate and nuts. We close on the rise of the “scholarly engineer,” and why understanding ideas may matter more than reading every line of code.</p><p><br></p><p>Demetrios Brinkmann: <a href="https://www.linkedin.com/in/dpbrinkm" target="_blank" rel="noopener noreferer">https://www.linkedin.com/in/dpbrinkm</a>Matt Carey: <a href="https://www.linkedin.com/in/mattzcarey" target="_blank" rel="noopener noreferer">https://www.linkedin.com/in/mattzcarey</a>Scott Crooks: <a href="https://www.linkedin.com/in/jshcrooks" target="_blank" rel="noopener noreferer">https://www.linkedin.com/in/jshcrooks</a></p><p><br></p><p>Timestamps:[00:00] Cold open[01:03] Meet Matt and Scott[01:20] What an AI gateway really is[02:28] Multi-region, multi-cloud model routing[05:00] Cloudflare’s agent stack and Cloudflare OS[06:43] The Flue framework from the Astro team[07:32] The best AI code reviewer[09:29] An engineering codex written for agents[11:48] Agent experience and evals[12:44] Miro’s Easel bot and three layers of review[15:55] Software factories and Astro hitting zero issues[18:58] Companies open sourcing internal tools[22:59] Hot takes[25:03] The scholarly engineer[28:50] Wrap up</p>
出典・この記事について
- 番組
- Agentic Conversations (formally mlops.community)(Demetrios)
- 元の回
- Multi-Cloud LLM Routing and Context-Aware Coding Agents
- 配信日
- 要約作成日
- 公式サイト
- mlops.community
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

Stripeはなぜ社内ドキュメント50万件を5,000件に絞ったのか
Agentic Conversations (formally mlops.community)

私のAIエージェントが車を買ってくれた:Hermes Agentと自己改善の最前線
Agentic Conversations (formally mlops.community)

スタンフォードが教えるAIネイティブ・エンジニアリング — Mihail Eric(Monaco)
Latent Space: The AI Engineer Podcast

1つのAIコストを削ると2つ生えてくる:Prosusが語る社内AI運用の総保有コスト
Agentic Conversations (formally mlops.community)

Kerr's Law の誕生:エージェント時代のボトルネックはチームの「メンタルモデルの同期」にある(Changelog Interviews #689)
Changelog Master Feed

エージェントにも OAuth2 が必要だ
The Stack Overflow Podcast