Anthropic Claude Code Team Automates Software Development Via Multi-Agent Loops

Anthropic Claude Code 团队通过多智能体循环实现软件开发自动化 AI 辅助软件工程的瓶颈已从编写代码转变为人类的工程品味与验证，这使得自主多智能体流水线成为提升开发速度的主要驱动力。 在这场直播分享中，Claude Code 负责人 Boris Cherny 与 Bun 创始人 Jarred Sumner 演示了他们如何利用自主 AI 智能体来管理、重现并修复 Bun 仓库中活跃的 GitHub issue。他们介绍了由 Claude 驱动的机器人“RoboBun”，该机器人能够自动重现 Bug、执行验证循环，并提交附带测试的 Pull Request。讨论内容涵盖了多智能体对抗性代码审查的机制、精心维护的 `CLAUDE.md` 文件对复合工程的关键作用，以及夜间扩展至数百个智能体如何将开发瓶颈从编写代码转移到工程品味和更高层级的验证。最终，他们通过在演讲过程中成功推送四个自动化 PR，展示了“自动模式”的迅猛发展。 Bun 代码库引入了由 Claude 驱动的自动化机器人 RoboBun，能够自动复现每一个 GitHub issue 并提交附带强制测试的 PR；在过去三个月中，它在主分支上的代码贡献量已超越其创始人 Jarred Sumner。 (2:08) 团队构建了结合 Code Rabbit 与 Claude 的“对抗性代码审查”回路，前者专精于代码风格与 CLAUDE.md 规范，后者深挖复杂架构边缘案例，大幅消除了开发者频繁切换上下文的成本。 (4:10) 维护详尽的 CLAUDE.md 文件成为多 Agent 工程化的核心基石，配合健全的自我验证闭环，支撑起开发者每晚自主并行运行数百个 Agent 的高扩展工作流。 (8:03) 完全自动化部署的瓶颈已从代码生成质量转向验证速度、正确性证明与回滚机制；此外，Claude Code 已展现出“爬山式”自主迭代能力，能通过研读系统底层代码持续优化库性能直至超越基准。 (14:25)

Bun creator Jarred Sumner and Claude Code lead Boris Cherny argue that software engineering is shifting from manual coding to managing autonomous multi-agent verification loops. However, they emphasize that human engineering taste and deep verification remain the primary bottlenecks preventing fully autonomous code deployment. Automated Issue Reproduction and Adversarial Code Review (2:08) In the Bun repository, every new GitHub issue triggers an automated Claude-powered bot named RoboBun that attempts to reproduce the bug and generate a pull request (PR). A strict verification gate prevents the bot from submitting a PR unless it proves a newly written test fails on the previous build and passes on the debug branch. Over the past three months, Sumner claims RoboBun has become a larger contributor to Bun's main branch than he is. Once a PR is opened, a multi-agent feedback loop—which Cherny terms "adversarial code review"—engages to refine the implementation. The Code Rabbit bot addresses stylistic consistency and project instructions, while a Claude-based review agent inspects the code for complex edge cases outside the immediate diff. Sumner estimates this AI-driven review is wrong only about 10% of the time, allowing multi-agent discussions spanning dozens of comments to resolve issues without developer intervention. Structuring Compound Engineering with CLAUDE.md (8:03) Sumner notes that scaling autonomous agents requires a meticulously documented environment, captured in Bun's `CLAUDE.md` file. This document specifies precise commands to build the compiled codebase and forward arguments, preventing models from executing stale binaries. It also details directory structures, test placements, and rules to print critical error diagnostics before less informative conditions. Cherny refers to this iterative process of updating project instructions based on model mistakes as "compound engineering." This structured feedback enables optimization workflows known in AI labs as "hill climbing." In one benchmark, Sumner instructed Claude to optimize Bun's new image processing library against `sharp` on a Linux environment. By directing the model to read JavaScriptCore code and eliminate unnecessary type array cloning, Claude autonomously iterated until the implementation beat the performance benchmark. Autonomous CLI Workflows and Parallel Execution (11:01) Cherny and Sumner report that development has evolved into a massively parallel, asynchronous operation, with Cherny running hundreds of agents nightly. Sumner relies on Claude Code's CLI in "auto mode," which bypasses continuous manual permission prompts so long-running tasks can proceed autonomously for hours. While an earlier setting completely skipped permissions, auto mode balances uninterrupted progress with safety against unintended file deletions. To optimize the CLI experience, Anthropic introduced a rewritten renderer activated by `CLAUDE_CODE_NO_FLICKER=1`, incorporating virtualized scrolling and terminal mouse interaction to maintain constant CPU and memory usage. With Claude executing terminal commands and sleeping between polling cycles, maintainers can multitask across several projects while agents resolve dozens of comment threads and submit multiple PRs within single sessions. Shifting Bottlenecks and Human Engineering Taste (25:01) Sumner observes that as code generation and basic test verification become automated, the operational bottleneck moves toward higher-level system planning, proof of correctness, and rollback safety. While bots can rapidly triage bug reports and feature requests from social media into functional PRs within an hour, complete end-to-end automation remains constrained by product philosophy. Sumner stresses that deciding whether to incorporate complex features—such as bundling an image processing library or executing large-scale rewrites—still requires human engineering taste. At the same time, the frictionless generation of code has altered open-source social dynamics. Maintainers no longer feel interpersonal guilt when rejecting machine-generated PRs, which Sumner and Cherny argue counterintuitively raises the bar for what gets merged. Because there is no obligation to salvage wasted human effort, maintainers can reject imperfect contributions freely, focusing solely on optimal solutions. While multi-agent systems and auto mode have drastically accelerated issue triage and routine code generation, the path to fully closing the loop—automatically deploying fixes from issue intake without human review—remains unresolved. The primary barrier is not the model's ability to generate plausible code, but the development of rock-solid verification guarantees and human-level judgment over software architecture.

Anthropic 的 Boris Cherny 与 Bun 创始人 Jarred Sumner 通过现场演示阐述了如何利用 Claude Code 与多 Agent 自动化流水线维护复杂开源代码库，指出当前软件工程的瓶颈正从代码编写转向深层验证与工程品味判断。 RoboBun 自动化 Bug 复现与严格验证闭环 (2:08) Sumner 介绍，Bun 仓库利用基于 Claude 的自动化机器人 RoboBun 处理 GitHub 上的每项 issue，由其在人类介入前自动尝试复现漏洞并提交拉取请求（PR）。为确保修改的可靠性，该流程设置了硬性验证卡点：提交的 PR 必须附带自动化测试，且测试必须在旧版本上失败、在包含修复的新调试分支上通过。在最近三个月中，RoboBun 在主分支的贡献量已超过 Sumner 本人，这使得开发者的主要精力从排查与编写代码转变为审查修复方案是否达标。Cherny 补充说明，Bun 作为系统级 CLI 工具，无需依赖浏览器即可完成测试，在闭环验证中具备天然适配优势。 多 Agent 对抗式审查与 CLAUDE.md 复合工程 (4:10) 在代码审查环节，Bun 团队部署了针对不同职责的 Agent 协同机制。专门检查格式规范与项目配置的 Code Rabbit 与专注于深层逻辑的 Claude Code Review 在 PR 评论区相互交互，有时甚至能就单次提交自动往复讨论 30 余轮。Sumner 评估指出，Claude 的代码审查错误率仅约 10%，显著优于早期审查工具，能有效发现跨越局部 diff 的控制流边界漏洞。为支撑多 Agent 高效自治，团队在 CLAUDE.md 中详尽规范了构建命令、目录依赖结构、测试放置原则及构建日志诊断顺序。Sumner 将经验总结为“凡需重复叮嘱之事皆应写入文档”；Cherny 则将这种根据模型失误持续沉淀规则的迭代过程称作“复合工程”。 夜间规模化并发与指标导向的“爬山法”优化 (14:25) 随着以 Claude 3.5 Sonnet 为代表的推理能力成熟，Cherny 与 Sumner 目前均在夜间并发运行数百个 Agent 自主处理开发任务。Sumner 展示了如何利用 AI 推进大规模工程，包括新增 HTTP/2 与 HTTP/3 服务端、图像处理库集成以及针对 Rust 的系统级重写。在图像库开发中，Sumner 为 Claude 设定了在 Linux 基准测试中超越 sharp 性能的目标，并指导模型查阅 JavaScriptCore 源码以避免冗余的类型数组克隆，最终成功完成了性能优化。Cherny 指出，这种向模型提供明确量化指标与闭环验证环境、使其自主迭代直至达标的模式在 AI 实验室被称为“爬山法”（hill climbing），是当前极具潜力但尚未被充分利用的工作范式。 人情负担消退与工程品味成为核心瓶颈 (25:01) 在结合 CLI 的“自动模式”（auto mode）后，开发者只需投入约 10% 的宏观注意力，即可让模型在后台自主运行数十分钟至数小时并完成复杂的多步修复。更重要的是，代码生成门槛的降低打破了开源维护中的社交心理障碍：面对 AI 提交的不当 PR，维护者不会产生否定人类劳动的负疚感，可以毫无压力地直接关闭，这一转变反而反常地拔高了进入主分支的质量门槛。Sumner 认为，当前阻碍完全脱离人工干预、自动部署端到端修复的障碍已不再是代码生成质量，而是深层验证耗时、正确性证明的严密性以及回滚机制的健全程度，在更高层次上决定系统架构方向的“工程品味”目前依然依赖人类专家。 该演讲展示了自动化 Agent 与系统级代码库闭环验证结合所带来的生产力跃进，但发言者也坦承，在缺少绝对完备的自动化深层验证机制与自动化回滚保障的前提下，让 AI 自主决策并合入部署仍不成熟，人类的架构决策与工程品味在短期内依然不可替代。

[1:02] Jarred Sumner: this is classic Jared doing work during a talk. [8:03] Jarred Sumner: every time that you find yourself repeating something, it should probably go in `CLAUDE.md` [11:01] Jarred Sumner: Opus 47 [11:01] Jarred Sumner: 47 [22:15] Boris Cherny: no you guys can delete stuff if you do that. I don't I think I'm not supposed to recommend that. [25:01] Jarred Sumner: every time there's a new bottleneck, you have to kind of automate that bottleneck and then there's always some other bottleneck after. [23:59] Jarred Sumner: every time somebody submits an issue, it it tries to reproduce it. [25:01] Jarred Sumner: maybe like an hour later there's a PR [27:30] Jarred Sumner: you don't have to feel bad when it's like claude. [28:45] Jarred Sumner: in auto mode I can let like cloud runs like for hours and hours at a time. Like I run it almost every night.

[1:02] Please welcome to the stage head of Claude [music] Code of Anthropic, Boris Churnney, [1:07] and creator of fun at Anthropic, Jared Sumner. [1:22] [music] [1:23] Hello. [1:27] All right. So, this is a developer conference. We're going to be doing a little bit of talking, [1:31] but mostly we're just be like coding. So, this is for the developers in the room. [1:38] I'm going to start by talking a little bit about how uh [1:42] how bun uses cloud code to to build and and maintain bond and also kind of how our setup [1:47] uh works and because it's kind of a a slightly more advanced setup than uh what's common today. [1:55] Um but first I'm going to get a few agents running to just fix some GitHub issues. Uh [2:02] this is classic Jared doing work during a talk. [2:08] Um so so in buns repo uh every time somebody submits an issue uh we have a cloud bot [2:16] automatically run and try to reproduce the issue. So you can see this person has [2:21] uh this this side effects and this is like one of the most recent issues. Um [2:27] and and we can see that Robun which is our our our bot uh uh went and managed to reproduce the issue [2:34] uh and submitted a PR automatically. Um and and this PR is like it has uh it all these [2:45] PRs always have tests. Um it's one of the actual hard requirements before it can submit a PR. Um uh [2:52] and so so the challenge here is like does this code look correct? Um and one of the [2:59] things we do to to check that is does the test fail in the previous version of bond and pass [3:05] uh in the in this debug branch. And it the bot actually can't submit a PR without that [3:10] uh being the case. And so this is just to make sure you understand. So this is like every single [3:15] issue that goes up in the bun issue tracker, you have Robo Bun automatically try to reproduce it [3:20] before anyone looks at it. Yeah. And this saves a lot of time because we have so many open GitHub [3:25] issues. Um it it really moves the challenge from is from just fixing and debugging the issue to [3:32] is this the right thing to merge? Like is this the right fix? Um how how good is it? [3:38] Is it is this doing like a 100% of yours? Is it like 10%. We can go to the insights [3:43] uh uh and go to contributors and then if we go uh last three months uh uh and this is [3:50] specifically to main we can see that Robbun is now a bigger contributor to Bun than I am [laughter] [4:00] and that's with merging not all of its PRs for sure. You can see we have a lot [4:03] of PRs open right now. Um the the challenge is really how do we know can we merge the PR? [4:10] Um, and that's the tests. Um, and then the other thing that's really interesting about this is uh [4:16] uh we have automatic code review bots that that run. Um, and then they're going back [4:23] and forth. So like Code Rabbit leaves a comment and then Roboun leaves a comment [4:30] and then they go back and forth and they and Code Rabbit did the [laughter] [4:39] I love this. And it also marks the comments as resolved when it's done. Um, [4:47] and you you can see they actually went a lot like there's a lot of back and forth here. There's like [4:50] 30 comments or something. And so you're using like a combination of agents. So this is like [4:54] code review. This is like quad code review and then also code rabbit and like you're using [4:59] them together. Yeah. And I I think basically like code rabbit is is good for like kind of stylistic [5:04] issues and things that are like make sure that it follows the cloud MD and then the the cloud [5:10] code review is really good at here's this really subtle edge case that would have taken me like 30 [5:15] minutes of reading all the code and having all the context to to like figure out and and so so [5:22] it's really good at surfacing bugs that you you need the full context to really understand. Um, [5:30] and I think basically it's it's really hard to to actually have all this automation without [5:37] having code review that is in the loop with the with claude there replying or replying is like [5:44] very performative but like at like fixing um and that's also a big part of like what used to take [5:51] so much time uh when like why PRs would take so much time to merge is because you'd have to like [5:59] like check out the branch locally, fix a lint error, then run the llinter locally, [6:04] then push it back up. And there's all this switching cost that's constantly there. [6:08] And so when you So I think this is like an especially good use case for for LLMs because [6:15] otherwise like it just takes up so much time to to ship and and I guess like especially for the [6:21] bun codebase because it's like it's systems code. It's, you know, it's very easy to repro issue and [6:27] then see if the issue is fixed because this is kind of back to what we were talking about before [6:32] with like this kind of verification loop. It's all systems code. So, it's really like a test case [6:37] on a particular architecture and you can essentially like repro or verify anything. [6:41] Yeah, that's one of one thing that makes it easier in buttons codebase is because it's a a CLI tool. [6:46] um because uh we don't need to run a browser to test things, but you can also just like [6:52] have something set up to like take a screenshot or record a video or those sorts of things. [6:57] Um in Bun's case, we don't need that, at least not yet. Uh uh there's a couple things [7:03] we could do that for like we have some front end stuff that would be nice. Um [7:08] but yeah, I think this is like the the direction that I think is really interesting is because [7:13] it saves so much time. Um and this is not this is something that like this is specifically for bun [7:20] but the more generalizable thing when because most products are not open source um uh is like instead [7:27] of an issue maybe the starting point is like a customer support ticket. So like you could you [7:31] could imagine automatically uh h passing customer support tickets to a cloudbot to then go and try [7:39] to reproduce the issue and then submit a PR and then uh having code review go back and forth. Um, [7:45] and that's where I think for a lot of companies it becomes a lot more impactful because it just [7:51] saves so much developer time. We should think of some kind of name for this pattern. It's [7:55] like adversarial code review or or something like that. Yeah, I don't know. Um, but I do think like [8:03] there's also a few other things about this that's uh like the if you just do this then it doesn't [8:10] quite work. the very first step you need is to like make sure the development environment [8:14] is set up. Um like I I think this has been talked about before but like uh like cloud MD is uh [8:21] very important um because otherwise it's going to just submit PRs that don't quite make sense for [8:26] you to merge. Um so like we very much emphasize in buns codebase that it runs this special command [8:32] to do the build. Um and this both builds and runs the command. So it like forwards the arguments. Um [8:39] because that's also one confusing thing is like uh because bun has to be compiled uh you want to [8:45] make sure that it's running the actual changes and not like a debug build that's like stale. Um [8:52] we also go into a lot of detail about how to run tests, how to write tests, um where to put the [8:58] test, um uh and a lot of like here's how here's all the issues that we've run into previously. [9:06] Like basically the the pattern here is like every time that you find yourself repeating something [9:12] it should probably go in cloud MD because the the question now is like [9:17] how do you make it maintainable to have lots of claws running all the time and uh [9:24] and to do that it needs to be written down. It needs to be documented. Um so like a really small [9:31] detail is like we we check the we have it like to make sure that claude sees the error message uh we [9:38] make it print the air message before uh the like less informative uh conditions. So, so this is [9:44] sort of like you have quad write a test and then the test is bad or something about it doesn't work [9:50] and then you see this kind of repeated like once or twice and then you just tell Quad add it to the [9:53] quad MD so that every time in the future when you write a test you do it correctly the first time. [9:59] Yeah. So this is like uh like compound engineering. It's kind of this. [10:05] Um and then also it's helpful to like give it an overview of where all the folders are [10:09] like what all like how the code is laid out. Um, and like about dependencies. Um, uh, uh, [10:21] another thing I think is interesting is like making sure that it can read your CI errors [10:25] and like build logs. like you want to you want to set up the agent to be able to [10:30] be able to read the code like to do the full loop of [10:34] make of like writing the code testing the code works checking CI monitoring CI um and uh uh [10:43] reading all the errors so that way by the time it gets to a person everything is like set up [10:48] like the ideal is that you read the code and you you have very clear indications that you can be [10:53] high confidence to merge it And the only way for that to be true is if it is set up for success. [11:01] It's interesting. I remember when we when we first met, you were talking about like your vision of [11:04] like everyone being able to run hundreds of agents in parallel and how that would work. And I feel [11:10] like I didn't really get it at the time. Although now like every night I I'm I'm running like [11:15] hundreds of agents every single night. And I feel like now I'm finally there. But this is a thing [11:19] that you've been thinking about for a long time. So it feels like this is sort of like the setup [11:23] in order to be able to kind of scale up agents way more like you need the self-verification so [11:27] that agents can run autonomously. Yeah. Like this has gone through many iterations in buns codebase. [11:34] Um we previously just had a discord bot where I could just at mention the bot and it would spin [11:39] up a container. Um uh it didn't have like the CI stuff. It didn't have uh the code review stuff. [11:46] Um, and it's so much better now, especially with like Opus 47. Um, [11:52] uh, uh, it's all this stuff is getting so much better. Um, oh yeah, we can also check on [12:00] how, uh, how it's going. It looks like it created a PR, the first one. Um, uh, [12:08] and it wrote tests. Um, so maybe while we look at this, I'm curious for like just to [12:13] get a show of hands for people in the room, like as people think about their development process, [12:18] raise your hand if it looks something like this where you have like a bunch of terminal windows [12:21] or desktop tabs and you're kind of pasting in issues. Okay, so that's like maybe half of people. [12:27] And then what if it looks like Roboun or something more like that where it's like [12:30] closing the loop a little more? It's like the next level of abstraction starting to get there. [12:37] Yeah, I think it's like it's not surprising because I think model capabilities are just [12:40] getting there. Like I think 47 is the first model where it's really felt like it's able to do this. [12:45] And in the past maybe you could do it with like a bunch of scaffolding like you just [12:49] throw a bunch of tokens at it and it can kind of work. But now it's like efficient enough. [12:52] You can actually do this dayto-day. Yeah. Um let's see. So the first PR is there. Uh [13:01] let's see if it did any others. Uh okay. It did two PRs. Um uh [13:10] this looks very plausible. It's cool. And this like this before after is this like [13:16] do you tell it to do that or sometimes it does it? Um uh it's it's pretty good about knowing [13:23] like when it should do that like when it's like a string formatting thing. Mhm. Um uh it's also [13:30] it also kept like bun style of the label which is good because node style is slightly different. Um [13:40] let's see does this change look good? Um [13:48] yeah. Uh mostly what I'm thinking right now is uh it it's it did this. This is good because like [13:55] you don't want to write one bite at a time. You want to write uh [13:58] in in chunks um and then it used uh saturation uh to to but I don't like [14:05] that or like it shouldn't have to do that. Does anyone here actually know Zigg sort of [14:14] I feel looking at this code and and you can see it followed the patterns from the from the cloud MD [14:19] await using and then that that pattern of of uh reading all the pro like resolving all the [14:25] promises at the same time and so like what's your workflow like when you when you see something like [14:30] this are you usually going in and like commenting or are you just going to wait for like code review [14:35] to come in and and drop a comment? Um, usually it depends on that like how complicated it is. Um, [14:42] in this case, I'll usually I'll wait for like this one is actually pretty simple. I feel pretty high [14:46] confidence that if the test pass then uh I would probably merge this. Um but I still would wait for [14:55] uh the code review for for at least for the cloud code review one to run um uh [15:01] just in case because what I really like about that is it will find things that [15:06] like from from that aren't in the diff that are like from tracing the control flow. um [15:14] which is what you want when it's like a human reviewing it is like somebody who [15:17] has a lot of context who who can think what are all the edge cases that this might run into and [15:22] and the signal to noise ratio is pretty good. It's something maybe like 10% of the time it's wrong [15:29] like that's and and for like how that used to be like with other code review products that [15:33] we've tried it was like basically you had to ignore most of what it said. Um [15:40] it's pretty cool. How how long how long has something like this worked? Is it like a latest [15:43] model thing or like have you had like like Robo Bun or this kind of like automated repro automated [15:48] fixing like this whole pipeline? Like how how long has that been actually possible? We can probably [15:53] like see this in a chart somewhere. Um that's kind of a lot of commits but I think that's [15:59] that might not be on main. That might be that the the Rust thing. Um uh uh [16:08] yeah I heard bun is going to be written in Rust soon. Is that I don't know. [16:12] We're there's a I just have a quad running and we'll see what happens. Um uh [16:20] but you can see like the the the volume of commits there. It's like kind of lower and then it it's [16:26] it's definitely gotten up a lot and then it's gone up a lot. Uh now really the bottleneck is like [16:33] do I feel good about merging this? Am I confident that its changes are correct? And that's new [16:38] because it used to be like the code wasn't good enough. What do you think? What do you [16:42] think is left like like what's it going to take before well like is there like a missing tool [16:48] or like a missing model capability or model version or something for you to kind of feel [16:53] like Robban can fully close the loop. Issue comes in and then like fix goes out automatically. [16:58] I think it needs a little bit more uh it needs it takes a lot of time to verify the changes [17:04] are correct. M um this was kind of already true like when it's a person pushing up a PR. Um uh [17:13] but I think the the challenge is like how do we make it how do we make sure to communicate [17:21] sufficient proof that the changes are correct. Um or making it easier to like roll back things. [17:29] Um that's I think those are kind of the two directions. But I think like [17:35] for the majority of like simple issues, we should probably be pressing merge a lot more. And the [17:39] bottleneck now is actually like CI and like making sure that uh like like the and and having [17:49] like fully running the code like the the making sure all the test stuff works. M um [17:55] but I think it's like basically there for like and the large projects are still non-trivial [18:01] but also I've been doing some pretty large PRs lately with with cloud mostly with in the [18:07] not as much with Robo Bun but with like cloud code um like uh we recently added support for uh [18:14] a built-in image processing library to bond. I could probably pull up the PR. Um, [18:19] uh, and that was Claude. Um, uh, and also we did a bunch of follow-up PRs, too. Uh, yeah, [18:27] it's like it's interesting because I think like when I look at different people using quad code, [18:31] everyone is like at a kind of different level of like sophistication or kind of like adoption [18:35] of this. And I think for me the hardest thing is the model changes very often. So I have to [18:39] like constantly reune and kind of recalibrate to what it can do. And like as an engineer, it's hard [18:44] cuz it's like a very weird technology. It's like the first technology I've used that's like that. [18:48] And I I sort of feel like this is actually the way that you do it is ahead of how the quad code [18:53] team does it for quad code itself. And to me like that that like the way the quad code team does it [18:57] is actually very automated. But this is like even further ahead. This is like almost like [19:01] full liftoff like full fully closed loop. Like in the last two weeks we've added an HTTP3 server [19:08] to bun. Uh there's a PR for an HTTP2 server. There's uh fetch support for HTTP 3 and HTTP2. [19:16] Um there's this image processing API. There's the ongoing Rust rewrite which may not ship. Uh [19:22] uh uh that that's like the most ambitious one I've done so far. [19:26] Uh and I done is too strong of a word because it's very much not done. Uh uh so and so even [19:32] something like this like this is like a benchmark. So Claude ran this benchmark for you. Yeah, Claude [19:38] ran this benchmark. Uh I gave it like a this ran on like a separate like uh on like a Linux box. Um [19:45] uh yeah and it and I was like make it faster than sharp and that's basically what I did. Um [19:53] I gave it like a few ideas like oh you could try like read this code in JavaScript core to like [19:59] figure out the how to like avoid cloning the type array when it's not strictly necessary. Um, [20:05] but like it it then went and did it and figured it out. Um, uh, and like uh, yeah, it's pretty crazy [20:15] honestly because it was these this wouldn't none of this would have worked uh, several months ago. [20:23] Yeah. I I feel like like within like within anthropic like within an AI lab, [20:26] you call this kind of thing hill climbing. And this is this idea that like if you give the model [20:31] some sort of metric and then you give it a way to verify its result, you can just make it iterate [20:36] and keep going and keep going until it hits that metric. Um, and this is something like [20:41] 47 I think is uniquely good at. And I think it's something really underutilized because I [20:46] think it's the first model that's actually very good at that. And if you just give it a target, [20:50] you give it a way to like improve the performance and you give it a way to measure, it'll just like [20:53] keep going until until it's done if you let it go in auto mode. Yeah. And we can also see like uh [21:00] this is another case where like the code review comments was really really helpful because like [21:06] there was like in this PR there's like 100 comments or something uh [21:10] and it's just going and fixing everything. Um uh uh like it goes on for a while and [21:19] in the meantime you're just working on something else. Yeah, I was this was not like my the thing I [21:23] was 100% focused on. I was maybe like 10% focused on this. I was doing like five things at once. Um [21:30] uh uh and this definitely wasn't possible like 6 months ago. Um 3 months ago like this is like [21:40] very recent that this is doable. Okay. So how how are our sessions doing? Yeah. So we have [21:47] one PR there. There's almost another PR coming up it looks like pretty soon. Um uh [21:54] this one is should be the trickiest one. Uh [21:59] um it mostly looks good though. Um like the or like it looks plausible based on this [22:05] these changes. I wouldn't exactly do it this way, but I think that's we need like a better [22:10] more optimized way to do this because that's a lot of checks. Um and and looking looking at [22:15] your setup here, so you're using you mostly use CLI. Yeah. And um do you always use auto mode [22:22] for permissions? Auto? Yeah. And then before that I used dangerously skip permissions. Um [22:29] [laughter] [22:30] no you guys can delete stuff if you do that. I don't I think I'm not supposed to recommend that. [22:37] Um but I think it's just not fun to wait for Claude to like press approve because then you just [22:44] like go off and do something else and then it's just been sitting there. So that's why auto mode [22:48] is really good because it's actually like a real way to fix that instead of just like trusting. [22:55] And I I also noticed like the the input like the little composer it's stuck to the bottom of the [22:59] screen. So you're using no flicker mode. Yeah, I'm using no flicker. Honestly, I think we should just [23:03] like make that the default because it's so much better. Um like you can see I can scroll really [23:08] fast and like like you could scroll fast before but like sometimes there would be a flicker and [23:14] now there's not. Um, have have folks tried no flicker mode for CLA? [23:21] Yeah, a few people. So, it's like we launched it on April Fools. So, [23:26] you can think it [laughter] in hindsight it came across as a joke a little bit, but [23:30] if you actually do quad code, no flicker equals 1 quad. So, just like set that environment variable, [23:36] we totally rewrote the renderer that's running in the CLI. So, it's using virtualized scrolling, [23:42] virtualized selection. And so what this means is like constant memory usage, constant CPU usage, [23:47] and also some nice stuff like uh like if Jared types, he can actually like click around the [23:51] composer. And so you can actually click and mouse events work, which is pretty crazy for a terminal. [23:59] So I'm just also having it monitor the PR. Um, and you can see it's uh [24:04] it ran some commands and then it's going to go to sleep for 20 minutes and wake back up. Um, [24:09] 20 minutes is probably a little bit too long, but it's okay. Um [24:13] uh oops. And what's that? Is that like using like a loop or something? I think so. Uh uh. Uh [24:24] yeah. And then it's let's see how else is it doing. And then the other ones are still [24:29] uh apparently it fixed an extra bug as well. Uh uh. Okay. So we got Okay, so it's been what? It's [24:38] been like 20 minutes or so. 25 minutes and we got one. How many PRs have we gotten? Two. Uh, [24:45] three PRs. Three PRs. That's not bad. Yeah. And I think we'll get a fourth one [24:51] once it finishes running the tests. And then in the meantime, Robo is still running and [24:56] kind of like generating even even more PRs. Yeah. Every time somebody submits an issue, [25:01] it it tries to reproduce it. Yeah. I I kind of feel like every like the the way quad code makes [25:08] you think is every time there's a new bottleneck, you have to kind of automate that bottleneck and [25:13] then there's always some other bottleneck after and you kind of move on to that and like it [25:17] started like writing code was the bottleneck and now it's no longer the bottleneck and then like [25:21] verification and running tests that was like the bottleneck and it's no longer the bottleneck and [25:25] now there's like a deeper layer of verification maybe that maybe that's it. What do what do [25:30] you think are the bottlenecks remaining? It's definitely this deeper layer of verification. Um I [25:34] feel like the bottleneck after that is going to be like planning like what should we do, what should [25:40] we not do and what is the right way to fix this. Um and like ideally Claude would be smart enough [25:47] or like we could trust Claude enough to to merge the PRs by itself. Um and I think like in a in [25:55] certain projects you could pro probably do that um and just have that be automatic completely. [26:00] Um, I think Bun is not yet or like it's not yet there for Claude or sorry for for Bun. Um, [26:06] but I think it'd be really cool if if like we had the tooling uh for us to feel confident enough [26:13] to do that. So So like right now Robbo it doesn't like build features. It it doesn't do like feature [26:18] requests yet. It it that's true. Yeah, it doesn't do feature requests but it we do also use it [26:23] sometimes. We we so we can also uh uh appmentntion it in either discord or slack and it will like [26:31] try to implement the feature. Um uh so sometimes when people are like hey bun is missing this thing [26:37] um then I just at mention the bot and maybe like an hour later there's a PR. Um [26:45] a bunch of times I've somebody's like tweeted at me something like [26:49] this can you fix this bug or whatever and that's basically what I do. Um, [26:53] and then I reply with a link to the PR. Um, should we add a Robbo account on Twitter? Um uh [27:03] and I think like so like it can do feature requests, but I'm I'm hesitant for it to [27:07] implement literally everything anybody asked for in a GitHub issue cuz that's kind of a lot cuz in [27:14] some ways it's it's kind of crazy to put something like an image processing library inside a bun, [27:18] but like we talk about engineering taste and there's like an element of taste that goes [27:21] into that. you felt like that's a good idea and like we're not sure yet if Quad is at the point [27:27] where you would also think this is a good idea but you know at some point in the future it'll [27:30] get there maybe it's starting to get and I do think that like PRs become suggestions like [27:38] having uh like not merging PRs used to be like you feel bad if you don't merge like a co-orker's PR [27:45] uh because like they put work into that but you don't have to feel bad when it's like claude [27:51] Um uh so like if the PR is wrong or or or or for for or whatever reason then you can just [27:59] not merge it. And uh but it does mean that like the like the the bar for uh what you merge is like [28:11] should should it be there because I think that's also a difference with with when it's like people [28:15] because with people you want you don't want people to feel bad about their like lost work. [28:21] Um, so sort of in some ways it does actually end up raising the bar for what you decide to merge. [28:28] Yeah, it's interesting how like as the bottlenecks move, the dynamics change a little bit. [28:33] It's sort of like having to trust each other, having to trust people on the [28:36] team. This kind of changes a little bit. Now it's a little bit more about like do [28:39] we have the right automation and like do we trust automation like as as a group? Yeah. [28:45] So I think we're almost at time. Is there any like maybe one last thing we want to [28:49] we want to show people where we can kind of check in on kind of like the progress that we've made? [28:55] Um I don't think so. Uh uh yeah, it's it's still going one last [29:02] onto that fourth PR and it's going back and forth like found a bug and then fixed a bug. Uh [29:09] looks like it's about to submit the PR now. Okay, let's wait for this one. [29:16] Love to have one more. This is like the cool thing about auto mode. It's like in auto mode I can let [29:20] like cloud runs like for hours and hours at a time. Like I run it almost every night. I'll [29:25] just have a bunch of quads running in auto mode. And I was like before this it just didn't work [29:29] cuz it always got stuck at some kind of permission request. And that was crazy. [29:34] Like this entire thing was one prompt and and that just ran for 30 minutes. Yeah. This is all I said. [29:47] Okay. So, it's pushing at it's about to submit the PR. [29:57] That sounds like the right fix, too. This has been an issue that's been open for a long time. [30:04] Okay. And we got a PR [30:14] Yeah, we can go to this issue and we can see how many up votes it has. [30:18] 20. Yeah, it's kind of a lot. [30:23] Cool. Maybe we can pause there. Um, but to me, this is just like such a cool vision [30:30] of where engineering is going, I think, for everyone in this room. And, you know, [30:35] we're going to see this first. we're going to have to figure it out first and then everyone [30:38] else is going to have to figure this out. So, you know, like we were talking about this morning, [30:42] just excited to be on this journey together. And like you can see, we haven't figured everything [30:46] out yet. Um, but I think like the mode that we're in is just constantly experimenting, [30:51] constantly trying to to see what the next bottleneck is so we can solve it. [30:57] It's very exciting. This is so cool. Like this stuff. [31:02] [applause] Cool. All right. Should I like Yeah, you can weave it. [31:12] Come on, [31:13] bro.