Claude Code：智能体进化的一年

It's hard to remember what that was like. Like it is, it's so different than what we're doing today. Like now I just have like armies of agents that are doing stuff. Like, I'm prompting one agent, or I have like an agent that's like prompting agents that's prompting agents, and it's like a tree of like thousands of agents.²

But I think it's just like the most important idea when working on this stuff is like every single time Claude makes a mistake, I don't tell Claude to do it differently, I tell it to write it to the CLAUDE.md or to like make a skill or something to do it differently. And if you can do this, then Claude can just like run forever. And I think the other thing that we kind of realized is the verification is really important. Like we didn't realize that.²

I sort of feel like this is this thing that just like everyone misunderstands. Because whenever we talk about verification, people are thinking like unit tests or they're thinking like lint or like type check. These are the things that are obviously really easy to automate and these are the things that were already automated.

But actually when we talk about verification for agents, it's something slightly different. It's like, can the agent run the thing? It takes a little bit of mental work to figure out how exactly do you do this because it's often not straightforward. And I think that's like, that's one of the challenges.³

What this skill does is Claude actually spins up a local desktop app and it uses computer use to click around on it. And so when I add a new UX, it clicks around to invoke the new UX. It also tests edge cases, and when there's an issue it fixes it and rechecks.⁴

But I also just feel like it's also just becoming easier because it's like essentially Claude writes the code, and so what matters a little more is like what's the idea that you have. And I feel like if you're a person that has like the product context and the business context and you're thinking about the design and the user, you're just gonna come up with better ideas.⁵

Like our finance team runs in Claude Code, they do their projections there. Um, data science. Uh like if you talk with our data scientists, it's so cool. It's just like everyone just has Claude Codes up on their screens. Yeah. Um, I feel like it's, it's remarkably versatile for different roles.⁶

One that I'm super excited about is routines. There's one engineer on our team who launched voice mode across all of our products. And, um, he has this routine set up that just listens for every ticket that comes, every GitHub issue, every bug report about voice mode, and his Claude just picks it up, proactively puts up a fix, and then pings the PR to him.⁷

So I shipped this small feature, and there was like an edge case in it that I didn't see. And so someone filed a bug for it, and I was gonna get to the bug that night. And as my Claude was working, it said, wait a second, another Claude has already fixed this. And I was like, how is this possible? Like I've never talked to him about this feature before.⁷

There's always like another person's Claude that's working on it. It's like, yeah that's been one of the changes. I feel like we're um, a while ago we were trying to figure out like how to use routines, and I feel like just like the agent SDK was this first idea that we could use Claude Code programmatically. But I feel like at the beginning it just wasn't obvious how do we use it? What do we use it for?⁸

And I think routines are the first really obvious application. And um I don't know, like it just does like all the code review. It babysits like every PR. You remember back in the day you used to actually have to like respond to code review comments? You used to have to like fix CI? You used to have to rebase? Yeah. Like I haven't done that in a long time.⁸

Auto mode. Auto mode. It's the best.⁹

Instead of plan mode. Yeah, because the newer models they don't actually need like a planning step anymore. I think this was really important for like Opus 4 through maybe 4.5. Then I think starting with 4.6 and definitely with 4.7, it just doesn't need that planning step.⁹

And at the time that was kind of the best we had a year and a half ago because we didn't have, you know classifiers, the model was not as well aligned as it is today. So auto mode was just such a, it was such a big step up because actually you don't want to read most of these requests. Just routing it to a different model and having it check for security works so much better. Yeah.⁹

It's just human nature when you accept 99% of requests that your eyes just glaze over when you read it. And so actually we feel that auto mode is more safe than reading every single permission prompt because it means that you're only paying attention to the most important thing and not like being spammed a bunch of things that are just 99% yes.¹⁰

And internally, um to actually get auto mode out to our users, we needed to really trust it first. And so what we did was we collected thousands of transcripts of like an entire agent trajectory and a permission prompt, and had auto mode classify whether or not it was safe.¹¹

Yeah I mean it's like it's honestly like a weird approach. I feel like there's like all these features the last year where the first time someone pitched it I was like ah no no way, that's not gonna work. And I feel like over time I've just learned like I'm actually wrong like so often now, because like building on the model is so weird. Yeah.¹²

Um, yeah I love loop. How do you use it? I think for loop there's this transition that we went through like a year and a half ago where we were like all right, there's source code, but actually the thing an engineer should interact with, maybe it's not the source code, maybe it's the agent.¹³

And so we made this leap of like I don't write the source code, I talk to an agent and the agent writes the source code for me. And I think right now what's happening is we're making the next leap. I don't talk to an agent anymore. I talk to loop or I talk to a routine. And it prompts Claude for me.¹³

I'm gonna put on my business cat hat. I have this like favorite case study this is like a Harvard Business Review from the 90s. And they were talking about like computers are here, why are we not seeing the productivity benefits? And it's just this like amazing snapshot into like what it actually felt like at the time.¹⁴

And it turned out that to get the productivity benefits from computers, what you had to do isn't like you have your paper filing cabinet and your paper and pen business process and then there's like a computer on the side that does something.

Actually what you have to do is you throw out the filing cabinet, you have to throw out all your paper and all your pens, and you put a computer in the center. And everything has to run through the computer. It has to be at the center of every business process.¹⁴

And I feel like for us Claude is just at the center of everything. Whenever I have a question I ask Claude. Whenever I write code I use Claude. Whenever I need a code review Claude does it. Uh whenever I need a security review Claude does it. Whenever I need to fill out a form or something, Cowork does it.¹⁴

I think it's just really exciting. Like I feel like now I don't have to bug people anymore. And when I interact with people it's because it's like fun and I get to collaborate with them on stuff and we get to create something together. It's not that like I need them, I need something from them. Cause like Claude can actually do a lot of that stuff now.¹⁵

And I also feel like as an engineer I've just never had this much fun doing engineering because the tedious part I don't have to do like I'm just coming up with ideas, I'm talking to customers. And every idea I don't have a to-do list anymore like Claude just builds everything.¹⁵

Everyone's gonna be both. I feel pretty strongly that these roles are merging. Like when we look at our team, our product team all writes code, our devrel team all writes code, our design team all writes code.¹⁶

So I think right now AI really benefits people who have a lot of curiosity, have a lot of product taste, who love to have this like end-to-end ownership. And now a lot of people are running like hundreds of agents. What are the products that you think people should be adopting as they transition from single to multiple to hundreds?¹⁶

And the thing that I would not have expected six months ago is probably half my engineering now I do on my phone. So I just have like I have so many agents running that I just start from my phone. I use remote control, which is like amazing now.¹⁷

And sometimes I'm like talking to someone and we come up with a new idea. I'll just start an agent on the spot. I'll like talk to it with voice mode and just have it build something. And I don't even have to go back to my computer anymore.¹⁷

But then it would be like pretty late and I was like hmm maybe he just like left it here by accident. And then it happened again the next day, and happened again the next day. And I was like wait this is so weird cause you're landing PRs but your computer's right next to me. And I remember you responding and you're like yeah I'm coding from my couch.¹⁸

This is a thing you know people used to talk about prompt engineering, they used to talk about context engineering. This is sort of matching where the model was at the time. Back in the days of Sonnet 3.5 you had to prompt engineer, back in the days of Opus 4 you had to context engineer.

But with the models of today you don't do any of this. You give it the minimal possible system prompt, the minimal possible tools, and then you let the model figure it out. Like you just have to give the model some way to pull in the context. I think that's the most important thing. How do you think about it?¹⁹

I see things very similarly. I'm a context minimalist. So my general philosophy is tell the model only what it needs to know, and let it figure out the rest of it. Um I think when you give the model too much context it's kind of like you're micromanaging it. And sometimes the model knows a better way to get to the same outcome. And I personally prefer to give the model that freedom to do that.²⁰

There's all these different ways to use Claude now, but I feel like in a year it's going to be a totally new set of things. And it's going to be so surprising if it's still these same things. Cause I think like we're seeing these giant trends happening right now. Agents are running for longer, they're more autonomous.²¹

And I don't know what it's going to be, and I think in large part it's going to be up to the team to figure it out. And this is um this is why I'm like so happy we run the team the way that we do where everyone just comes up with ideas and everyone is able to think about the product, everyone talks to users all the time. Because I don't think these ideas are gonna come from us, it's gonna come from the team.²¹