The Exchanges

Q And, uh, how does the MCP, uh, part work? So MCP, the model context protocol being something that, that you guys at Anthropic, uh, defined, how does that work? You just leverage the, the protocol to connect to any tool?

A Yeah, exactly. Quad code is a MCP client and an MCP server. And what this means is if you give it tools to use, so maybe at your company, you have a bunch of MCP tools that you build for kind of all your systems to integrate with. Like I said, maybe there's one to integrate with Jira, maybe there's another one to read Slack, and maybe write messages to Slack. Maybe there's another one to, um, you know, fetch some internal knowledge base or something like this. You're going to plug this into a bunch of your tools. You can plug it into Quad AI, into Quad Desktop. You can also plug it into Quad Code. So it gets all the same tools that you do.

AI assessment note: “Quad code is a MCP client and an MCP server.”

Q Fascinating. And do you have to declaratively add to the memory, or whether today or in the future, the memory will automatically pull from the context and sort of improve itself?

A Yeah, you have to add to it manually today. We, we've actually had a bunch of internal experiments to do automatic memory so Quad can automatically remember things. The problem is there's kind of two ways in which it fails. One is that it remembers things that it shouldn't. So for example, if I say make the button blue, it might remember the user always wants the button to be blue. And this is, you know, maybe that's the case for this button. That's not the case for every button. And then sometimes it doesn't remember very important things that it should remember. And so for the last few months, we've been doing a lot of experiments to try to get this performance really good. And it's something we've been using internally. And at some point when we're happy with it, it's something we're, we're going to release for everyone. Um, but generally our bar is if we find ourselves really happy with it and we find ourselves using it every day, then we release it to everyone. And this one's not quite there yet.

AI assessment note: “Yeah, you have to add to it manually today.”

Q And then the, the key question for, uh, agents is always, uh, the, the level of autonomy, uh, versus, uh, human in the loop. How, how do you guys think about that and where do I get pinged as the human coder, uh, in my cloud code workflow?

A The default behavior is there's always a human in the loop. This is, this is super important because this is, in the, in the end, this is a model and it's not predictable and, You want to make sure that it doesn't do anything dangerous. Um, so yeah, there's, there's always a human loop. So for actions that we know can't have any, um, kind of dangerous repercussions. So for example, reading a file, we know this is inherently safe. We just let the model do this in the folder that you let it do this in. But for other actions like, uh, editing a file or running a command or using the internet, this always needs a human in the loop and it always needs a human to approve it. There's ways to reduce this burden a little bit. So, for example, if you find yourself always approving edits to the same file, or always approving the same command, there's a, there's a settings file that you can configure across your team, and you can use this to essentially allow list or block list certain commands, or certain files that you always want the model to be able to edit without human approval, or you never want it to be able to run.

AI assessment note: “editing a file or running a command or using the internet, this always needs a human”

Q So in the, in the same vein, what do you find yourself, uh, using cloud code for as a, as, as a leader and engineer, what's your, uh, you know, daily use case?

A Yeah, I use it all day for all, all sorts of stuff. The, so obviously, like I said, code-based research, If I'm working on a piece of code I'm not familiar with, I'll just start by asking quad code to tell me about it. Whenever I'm working on a small feature, I'll usually use quad code in GitHub Actions. So I'll just say add quad, I'll make a new GitHub issue, and then I'll say add quad, implement this feature for me. And it'll just do it usually in one shot. Um, and sometimes I'll do this on the command line too. So I'll just say, you know, implement this feature and make a pull request and, you know, I'll come back a few minutes later and it's done. Then there's this kind of other work where it's a little bit more complex. You can't really do it in one shot. It's, it's not as simple as changing a piece of text or changing a button or building a small feature. Maybe it's like something more involved. There's probably two workflows I have here. One is for really complex stuff. I'll prototype it a bunch. And this is something that I did even before a quad code. You know, when you, when you write a complex piece of code or a complex feature, often engineers will write it a few times. Because you don't actually know the right way to do it, and so you'll try one approach, you'll try a second approach, you'll try a third approach, and you'll kind of figure out the edge cases for eac…

AI assessment note: “Yeah, I use it all day for all, all sorts of stuff.”

Q Okay, great. So that's, uh, the actions part of the agentic workflow. Uh, let's talk about the, uh, awareness and memory of this. One of the exciting features is that, uh, Cloud Code can, can connect with the existing sort of code knowledge, uh, in, uh, the company. How does that work?

A There's a few different ways to pull in context and this kind of knowledge from, from the company. The simplest one is just looking at files. There was this, ah, there's a, there's a few different approaches actually to reading files. So I'll, I'll go a little bit into depth into the way that actually happens. In the past, the thing that people use the most is this thing called RAG, and essentially this is a technique where you take the whole code base, and this actually works for any document, set of documents, it's not necessarily code, but you take a set of documents like, uh, like all the files in the code base, you do this kind of indexing step, and then you store essentially this database of all the knowledge that's in these files in a very, very particular form that makes it really easy for the model to search. There's a lot of trade-offs to doing this. The indexing takes time. It's pretty expensive to maintain this database. It's quite tricky practically to make sure that security is really good and privacy is really good because, uh, it's just a, it's, it's a very sensitive information like your code base. And so you want to keep it really safe. And so quad code actually doesn't use this technique called rag. Instead, what it does is it just searches files the same way that a human would. You can think of it like, um, you know, at the engineering level, it uses the too…

AI assessment note: “Instead, what it does is it just searches files the same way that a human would.”

Q a key part, uh, of the product is the agentic aspect, as, as you described. Agentic, again, in an effort to, you know, go into definitions and, and, and make this, uh, interesting for, for everyone. Argentic is one of those terms that, um, you know, everybody uses, but it's sort of, um, you know, unclear what that actually means. What does Argentic mean in the context of Cloud Code?

A Yeah, when you think about the ways that, uh, LLMs work and the way you interact with them, there's the old kind of LLM, which is you send them a message and they send you a message back, and this is sort of these chat applications that everyone knows and uses all the time. There's a newer kind of, uh, application, a way to interact with LLMs, where you send them a message and they'll send you a message back, and then They might do a little bit more. So we call this tool use, uh, is, is one of the things that they might do if you give them tools. So for example, a tool might be read a file or search something on the internet or edit a file or something like this, then they'll use tools to answer your question. And so for example, if you ask, uh, you know, what, what's the weather today with the old style LLM, uh, interaction, it'll just kind of use its existing knowledge and its existing training to try to answer that query. But if it's agentic, what it might do is it'll say, okay, I'm going to look up the weather, and then maybe if it has some kind of tool to check the weather, it'll reach out to the internet or wherever that tool is, it'll, it'll use that tool, it'll get the response back, and then it'll answer the question. And this tool use, this is kind of the essence of being an agent because without tools, it's very different. It's very difficult to be agentic. You could…

AI assessment note: “This tool use, this is kind of the essence of being an agent”

Q Yeah, it's a, it's a, there's a concept of, uh, context pollution, right? That's, that's one of the way, uh, people describe it, right? And then, uh, Cloud Code can handle both very precise tasks, like debugging, or much broader tasks, uh, like a, a broad refactor, for example. Is the idea that, uh, the broader the task, the more sub-agents you would have in the current context?

A Yeah, that's, that's probably one way to think about it. Generally, when, uh, when we introduce QuadCode to new people, we actually suggest, like you said, QuadCode can do everything, and this is one of the things that makes it a little bit hard to use if you're an engineer that's used to, you know, essentially like text completions in an ID. It's a very different kind of AI coding experience. And so generally the thing that we recommend is start with something simple, like, uh, just ask questions about the code base. So don't even code. Don't use any tools. Um, just, just ask the model questions. You know, how, what does this file do? Um, where is the file that does this thing? If I want to make, uh, a new whatever, um, how do, how do I do that? So just ask it questions like that. And for this generally the main model can do it and you don't really need sub-agents. But then as you get a little bit more sophisticated You might want to start putting up the work. So if you ask the model maybe to make a small change, like, like I said, like make the button red or make the button blue, you probably don't need subagents. But if you do something a little bit fancier, like build a new section of the website that does blah, blah, blah. Then you might want to have a few sub-agents. Maybe one is the software architect, and it's responsible for planning out the work. Another one is maybe …

AI assessment note: “Yeah, that's, that's probably one way to think about it.”

Q million in annualized revenue. Uh, that was That's at least what was reported. I'm not going to ask you to, to confirm or deny, but that's what the press is saying. Perhaps even more importantly, there are rave reviews everywhere and gushing videos calling CloudCode the best coding agent by far. So amazing launch of the product. Did you have any sense that it was going to be this successful?

A Not really. We started by building something that's useful for us and, um, I built a thing that, you know, first and foremost was just useful for myself, and it's, uh, it's something that I found myself using every day, and then I gave it to the team, and I saw that the team started using it every day, and pretty soon after we launched it, most of Anthropic was a daily active user, and so I think at that point we had kind of a hunch internally that maybe we had a hit on our hands, but it's still not obvious, because It's in a terminal. It's kind of a weird form factor. Everyone is coding in IDEs. Are people going to like this? Is it going to be that useful? Can you actually use it for a lot of coding? We, we had no idea.

AI assessment note: “Not really. We started by building something that's useful for us”

Q Okay, great. So what did you guys choose, uh, to operate at the, uh, terminal slash CLI level?

A Honestly, it was sort of an accident. We were thinking about what can we build in the coding space? It seems like the models are getting a lot better at coding, so maybe there's something that could be built here. And at the time, like I said, I was prototyping some of these ideas. Some of them were kind of coding adjacent. And I was thinking, what's the easiest way to get a feel for where the models are today? Because there was this feeling where the models can do so much But no one's yet built the product that can harness this capability. And in the AI world, we call this idea product overhang, where the model is just capable of all of these things. And there isn't really yet a product that can kind of capture this, uh, and, and, and let a person use it. And I remember there was this moment back in, uh, yeah, sometime late last year, the other engineer on the team, Sid and I were in a room and we're whiteboarding and we're like, okay, what do we build? And just in 15 minutes, we kind of throw up a few ideas on the whiteboard and we can do a CLI, we can do Some sort of IDE extension. We can do something based on the web. And we kind of closed our eyes and just picked one. And it kind of seemed like CLI is the simplest. And generally that's the way that I approach product, which is start with a simple thing first. Um, so this kind of made sense. In hindsight, there was a lot of…

AI assessment note: “And it kind of seemed like CLI is the simplest.”

Q What are some of the actions if, if taking an action is, uh, one of the core, uh, part of, of being an agent, what kind of actions can, uh, Cloud Code take?

A Cloud Code can do pretty much anything that a person can do on their computer. There isn't really a limitation besides, uh, safety. Um, this is something we think about a lot, uh, to kind of place intelligent limits on what it can do and put a human in the loop at the right points. To make sure that if action is potentially dangerous or destructive in any way, uh, that a human has to prove it first. But besides that, the model can do pretty much anything. So, you know, reading files, writing files, uh, running commands on the system, editing things. Uh, it can reach out to the internet. Um, you know, obviously most of these, again, with, with human approval. Um, and then there, there's ways to customize it however you want. So if you have a bunch of MCP tools, for example, to read, uh, your JIRA issue tracker, or to open a browser, or To open an iOS simulator, the model can use these two.

AI assessment note: “reading files, writing files, uh, running commands on the system, editing things.”

Q As a last theme for this conversation, the sort of the elephant in the room is what that means for coding and coding as a Profession. What's your general sense for what coding is going to look like in a few years from now?

A Yeah, it's a little bit hard to say. Um, you know, a few years from now is, you know, in AI time is like decades, you know, in, in normal time. I think even today for a lot of professional coders, it's really easy to lament the state of coding and to think, you know, I used to write this code by hand and, you know, now it's this agent doing all of it. I think actually being the one that does this work, it's incredibly exciting to have an agent write the code. It feels very empowering as an engineer because I can explore a lot more ideas than I could before. I can do it much faster. I can work in domains that I know literally nothing about. You know, maybe I don't know iOS, but I can write an app. Um, because I can just generally code review the code, and I can see it looks reasonable, but Quad actually does all the writing and all the, all the testing of it. There's, um, one engineer on the team, Lina, she still writes C++ on the weekends, sometimes by hand, she was telling me, because, um, you know, like as a programmer, this is kind of one of the things that we enjoy, because sometimes you have to get down to the metal, and you have to kind of do it this way. But I see this as a transition the same way that in the sixties, there was this transition between punch cards and assembly, and then later on between assembly and Fortran and COBOL and the first high level languages. An…

AI assessment note: “programming is no longer direct text manipulation”

Q in the information that, um, Anthropic has a little above 40% of the market for code generation, while OpenAI has 21%. So, uh, clearly, Anthropic is, uh, powering its way to win this, this market. Why is that? Is there something about the way the model is trained, the data is trained on, or the focus of the, of the training that makes it so great for coding use cases?

A I think Anthropic has a lot of really great researchers and, and coders, and for us, this is kind of a natural way to think about the kinds of abilities the model should have, because you think about what should the model be able to do, and the, you know, as a programmer, the first thing you think is, oh, it should be able to do the thing that I'm doing, um, and it should be able to help with me, and I can pair with it like I would, uh, like another, another engineer. So I think just at all, at all levels and across the company, this is a way that we think about it. There's also something about it where maybe coding is the way that we get to the next level of intelligence. If you call it like AGI or ASI or whatever, the model needs some way to interact with the world and for a model, the natural way is code. And so from, you know, the mission of the company is to build a safe artificial intelligence for everyone and safe ASI. The way the model interacts is through code. And so this is the thing that we should start to learn about now. And this is also the reason that, one of the reasons that we release cloud code is just to learn how people use it, and to learn how to make this thing safe, to learn how it behaves in the wild, so that we know what to do next.

AI assessment note: “coding is the way that we get to the next level of intelligence”

Q goal of being, you know, as lean and, and lightweight of an interface on top of the model. So we just talked about how you can approve, uh, actions that the agent takes. Um, what about, what, what about the, The, the, the rest, does it basically feel like you'd be using a regular CLI or is it something, um, you know, different that one needs to get used to?

A We tried really hard to make quad code something really beautiful that everyone feels is something that we put a lot of care into because we did. And I think when you use quad code, you can feel that this is something that we use every day. You know, at this point, most code at Anthropic is written using quad code and almost everyone at Anthropic is using it every day. And when we look at customers that start to use quad code, they, they use it kind of more and more and more. And with a product like this, you want it to feel really smooth and really beautiful and something you really enjoy using. And that's been really fun from an engineering and design point of view, because we built in a terminal, like I said, terminals have been around for, you know, 50 plus years at this point, and it really feels like we're rediscovering how to design for a terminal. Um, because since terminals were first invented, the, the design world moved to, uh, moved to web, and then it moved to apps, and there's kind of different design principles you can take here, and we try to apply these back to cloud code, even though it's running in a terminal. And there's a lot of details that we spent a lot of time on, like the way that we represent statuses for every item with this kind of blinking dot that turns red or green to indicate whether it succeeded or failed. The, even the loading indicator, like …

AI assessment note: “it really feels like we're rediscovering how to design for a terminal”

Q Okay, great. Let's start from, uh, the top end for anyone that, uh, is just starting to learn about cloud code. How do you describe it?

A Cloud Code is a agenda coding tool. Uh, the way to think about it is, uh, when engineers program, there's kind of different ways to program, and this has changed over time. This is, this has changed a lot over time. There was, you know, 5070 years ago, the way that people programmed looked very different than the way it does today. So, you know, 70 years ago, if you talked to a programmer, they would have been like, oh yeah, I take my punch cards, and I punch holes in it, and I put in this IBM kind of typewriter-y thing, and it punches holes, and that's programming, and then you feed this into this mainframe. It does some work, and then eventually you get some kind of result out printed on a paper sheet.

AI assessment note: “5070 years ago, the way that people programmed looked very different”