Anthropic 如何每天合并 3 万行代码变更：AI 时
字幕摘录
| 时间 | 英文 | 中文 |
|---|---|---|
| 0:12 | I'm pretty excited to be here and talk about who verifies the software factory. | 我非常兴奋地来到这里 谈论谁验证软件厂。 |
| 0:18 | Before we get started, can you give me a quick show of hands? | 在我们开始之前 你能给我举手吗? |
| 0:22 | Raise your hand if your coding agent has opened any pull requests in the last one week. | 如果你的编码代理 已经打开任何拉拉请求 在过去一周。 |
| 0:28 | Almost everyone. | 几乎是每个人。 |
| 0:29 | Thanks. | 谢谢 |
| 0:30 | And keep your hands raised up if you have merged pull requests and reviewed every line | 如果你把拉动请求合并了 并且检查了每一行 |
| 0:36 | of code that was merged. | 代码被合并。 |
| 0:39 | Oh, I do see a few people. | 哦,我确实看到几个人。 |
| 0:40 | That's excellent. | 这很好 |
| 0:42 | So what I'm going to talk about is the gap between the two camps, merging requests without | 所以我要说的是 两个阵营之间的空隙 合并请求而不 |
展开字幕全文（324 条）
| 序号 | 英文 | 中文 |
|---|---|---|
| 1 | I'm pretty excited to be here and talk about who verifies the software factory. | 我非常兴奋地来到这里 谈论谁验证软件厂。 |
| 2 | Before we get started, can you give me a quick show of hands? | 在我们开始之前 你能给我举手吗? |
| 3 | Raise your hand if your coding agent has opened any pull requests in the last one week. | 如果你的编码代理 已经打开任何拉拉请求 在过去一周。 |
| 4 | Almost everyone. | 几乎是每个人。 |
| 5 | Thanks. | 谢谢 |
| 6 | And keep your hands raised up if you have merged pull requests and reviewed every line | 如果你把拉动请求合并了 并且检查了每一行 |
| 7 | of code that was merged. | 代码被合并。 |
| 8 | Oh, I do see a few people. | 哦,我确实看到几个人。 |
| 9 | That's excellent. | 这很好 |
| 10 | So what I'm going to talk about is the gap between the two camps, merging requests without | 所以我要说的是 两个阵营之间的空隙 合并请求而不 |
| 11 | reviewing manually and merging requests after doing some reviews. | 进行一些审查后,手动审查并合并请求。 |
| 12 | So my name is Manish Kapoor, and I lead the product and developer relations team at Sonar. | 我叫Manish Kapoor, 我领导着Sonnar公司的产品和开发商的关系团队。 |
| 13 | And I'm excited to be here with my guest speaker, Adi Osmani from Anthropic. | 我很兴奋能和我的客座演讲者, 来自Anthropic的Adi Osmani在一起。 |
| 14 | Adi has been at Google for about 14 years. | 阿迪在Google工作了大约14年. |
| 15 | Most of you probably know him, and he's written lots and lots of content and articles. | 你们大多都认识他 他写了很多内容和文章 |
| 16 | I learn a lot from Adi, so I'm really fortunate to have him here stand next to me and share | 我从阿迪那里学到了很多东西,所以我真的很幸运,有他站在我旁边分享 |
| 17 | his wisdom with all of us here today. | 他今天在座各位的智慧。 |
| 18 | 他最近加入了Anthropic。 | |
| 19 | And Adi, so congratulations on your role at Anthropic. | 亚迪,恭喜你在安特罗皮克的角色 |
| 20 | Would you like to tell the team what you're working on at Anthropic, what your role is? | 你想告诉团队你在Anthropic的工作, 你的角色是什么? |
| 21 | Yeah, sure. | 当然 |
| 22 | So I'm responsible for making sure that you all have a really great developer experience | 所以,我负责确保 你们都有 一个非常好的开发经验 |
| 23 | building with quad code. | 以四进制码构建。 |
| 24 | And any time that we discover there are opportunities to make it even smoother, we bring that right | 每当我们发现有 机会使它更平滑, 我们带来的权利 |
| 25 | back into the product. | 回到产品。 |
| 26 | We ship a couple of times every single day. | 我们每天发货几次 |
| 27 | And so we're very interested in trying to make sure that you have the best possible | 所以,我们非常有兴趣 试图确保你得到最好的 |
| 28 | experience as often as possible. | 尽可能频繁地体验 |
| 29 | Cool. | 凉爽。 |
| 30 | Thank you. | 谢谢 |
| 31 | All right. | 好吧,我们走。 |
| 32 | So let's start with some problem and context setting slide. | 让我们从一些问题和背景设置幻灯片开始。 |
| 33 | There are two numbers on the slide. | 幻灯片上有两个数字。 |
| 34 | CircleCI今年分析了2800万CI工作流程. | |
| 35 | And they noticed that the throughput jumped by about 59%. | 他们注意到吞吐量猛增了约59%. |
| 36 | But yet, 3 in 10 attempts to merge that code into production are failing. | 然而,10分之3试图将密码合并到生产中却失败了。 |
| 37 | And this is a stat from one, independent stat from CircleCI. | 这是来自CircleCI的独立数据 |
| 38 | And I think I've seen the same trend for most of our customers that I speak to every day. | 我想,我看到了同样的趋势 对于我们的大部分客户 我每天说话。 |
| 39 | This is a quote from Adi. | 这是阿迪的引言 |
| 40 | A software factory is harnessing loops at scale. | 一个软件厂正在大规模利用环路。 |
| 41 | This resonated quite a bit for me. | 这对我来说很有共鸣 |
| 42 | And if you've not read that article, you can look it up. | 如果你没读过那篇文章 你可以查查 |
| 43 | It's pretty educational article and talking about future, how the future will look like | 这是相当教育性的文章 谈论未来 未来会是什么样子 |
| 44 | in terms of software factories. | 在软件工厂方面。 |
| 45 | So Adi, would you like to expand on what you were thinking when you wrote this article | 所以阿迪,你想多说说 你写这篇文章时的想法吗? |
| 46 | about soft loops and harnessing the loops at scale? | 关于软循环 和利用循环的规模? |
| 47 | Yeah, absolutely. | 当然 |
| 48 | So a lot of folks here have probably heard this term software factories. | 因此,这里的很多人可能听说过这个术语软件工厂。 |
| 49 | And there's two types of factories that are at least worth keeping in mind. | 而且有两种工厂至少值得记住. |
| 50 | Software factories harnessing loops at scale. | 软件厂在规模上利用环路. |
| 51 | A lights-on factory is one where you have agents that are helping you with planning, | 一个灯光工厂 就是有特工协助你计划的地方 |
| 52 | building, checking, and shipping code again and again. | 建筑,检查, 和航运代码一次又一次。 |
| 53 | But you, as a person, still apply your judgment and taste inside the picture. | 但你,作为一个人, 仍然运用你的判断和品味 在画面中。 |
| 54 | And then we have a dark software factory, which is a little bit closer to how a car | 然后我们有一个黑暗的软件厂, 这有点接近汽车 |
| 55 | factory with robotics might be working, where it's constantly doing the work. | 有机器人的工厂可能正在工作, 在那里它一直在做工作。 |
| 56 | So it's scoping, building, checking, and you, as the person, move down further to thinking | 所以,它的范围,建设,检查, 而你,作为个人, 走下去思考 |
| 57 | about intent and you're pulled in where they're really the greatest of exceptions to the case. | 关于意图,你被拉进去 在那里他们真的是最大的例外案件。 |
| 58 | But that's sort of what I was thinking. | 但我也这么想 |
| 59 | I think a lot of teams are somewhere on this spectrum and are on this journey to figure | 我想很多团队 都在这个谱系上 并正在这个旅程中 以图图 |
| 60 | out how do I get further and further into automation in a way that I can trust it. | 我如何一步一步地进入自动化 以我能够信任的方式 |
| 61 | So AI coding is now basically a solved problem. | 所以AI编码现在基本上是一个解决的问题. |
| 62 | We all know agents are there. | 我们都知道特工在那里 |
| 63 | 85% of the organizations are already using coding agents. | 85%的组织已经在使用编码代理。 |
| 64 | It'll go to 95% and very soon to 100%. | 会达到95%,很快达到100%. |
| 65 | 76% of the developers that generate AI code, they are saying that code is being produced | 76%的开发者 生成AI代码, 他们说代码正在生成 |
| 66 | at a speed that it's impossible for them to verify it. | 以他们无法证实的速度 |
| 67 | So they are spending more and more time reviewing the code. | 因此,他们花更多的时间 检查代码。 |
| 68 | So what do you think is the most pressing question that engineering leaders have right now? | 那么,你认为什么是 工程领导人现在最迫切的问题? |
| 69 | So I think that teams have gotten better working with agents over the last year. | 所以,我认为球队得到了更好的工作 与特工去年。 |
| 70 | Models, harnesses have also gotten better. | 模特儿,吊带也变好了 |
| 71 | And what I see a lot of leaders being interested in is how do we enable you to ship code faster | 我所看到的很多领导人 感兴趣的是 我们如何让你更快地运送密码 |
| 72 | with confidence? | 有自信吗? |
| 73 | And that confidence piece is very important. | 而这个自信是非常重要的。 |
| 74 | Agents can now generate ways for you to check your code much more cheaply than ever before. | 特工们现在可以创造出方法 让你比以前更便宜地检查你的密码 |
| 75 | And if you think about the things that enabled us to call good code high quality, those were | 如果你想想那些能让我们 称好代码质量高的东西 |
| 76 | things like good tests, benchmarks, specs, anything that gives us confidence. | 诸如良好的测试 基准 规格 任何能让我们有信心的东西 |
| 77 | So agents are able to help us generate that a little bit more easily these days, a little | 因此,特工们可以帮助我们 生成一点点容易 这些年,一点点 |
| 78 | bit more quickly. | 再快一点 |
| 79 | We can, of course, refine those things. | 当然,我们可以改进这些东西。 |
| 80 | But I think that those checks are one of the most important things we should be spending | 但我认为这些支票 是最重要的事情之一 我们应该花费 |
| 81 | our time on these days. | 这些天我们的时间。 |
| 82 | If you are hoping to get the best quality output from your agents, you need to give | 如果你希望从你的经纪人那里得到最好的输出, 你需要给 |
| 83 | your agents a way to verify that the work is correct. | 您的代理人可以验证工作是否正确。 |
| 84 | If you've got that in place and you can make sure that the code is both correct and complete, | 如果你有这个,你可以确保代码是正确和完整的, |
| 85 | that can easily two to three by the quality of your outcomes. | 以你的结果的质量来说,这很容易做到2到3。 |
| 86 | Yep, yep. | 对,对。 |
| 87 | Absolutely right spot on. | 绝对是正确的位置。 |
| 88 | I think code confidence, we have been running an ad, code confidence is scarce and you really | 我认为代码信心,我们一直在运行一个广告,代码自信是缺乏的,你真的 |
| 89 | need to build it up now that the coding problem is solved. | 在编码问题解决后,需要逐步建立. |
| 90 | So actually, yesterday night, so this might be a surprise question for you, yesterday | 所以,实际上,昨天晚上, 所以这可能是一个惊喜 问你,昨天 |
| 91 | night I read a post from Anthropic and I encourage all of you to read it. | 我鼓励你们所有人阅读。 |
| 92 | 它讲述了Anthropic的运行规模. | |
| 93 | They are merging about 30,000 code changes every single day. | 他们每天都在合并大约30,000个代码更改。 |
| 94 | They used to do about 3,000 or so, 2,500 or so code changes just a few months back. | 他们以前大约做3000左右,2500左右的代码在几个月前就改变了. |
| 95 | They have scaled from 2,500 code changes to 30,000 code changes per day. | 它们已经从每天2500个代码更改缩到3万个代码更改. |
| 96 | That is mind boggling. | 那是心灵的摇摆。 |
| 97 | So what does code review look like at Anthropic and with the Cloud Code team that you're working | 因此,什么代码审查 看起来像Anthropic 和云代码团队 你的工作 |
| 98 | with? | 和谁? |
| 99 | Yeah, these are great questions. | 是啊,这些都是伟大的问题。 |
| 100 | So we were merging about 2,500 changes all the way back in May and even for our recent | 所以我们在5月份 合并了大约2500个变化 甚至对我们最近 |
| 101 | refresh of Cloud.ai, we put out a case study saying that we kind of three by the performance | Cloud.ai的新鲜度,我们做了个案例研究 说我们有三分 |
| 102 | of it. | 对 |
| 103 | That was accomplished by merging about 3,000 pull requests in a very short space of time | 这是在很短的时间内 合并了约3000个拉拉请求 |
| 104 | with engineers, product designers, a number of different functions contributing to it. | 与工程师,产品设计师,一些不同的功能对此作出贡献. |
| 105 | So our workflow for being able to get verification and code review in place is that we first | 因此,我们的工作流程 能够得到核查和代码审查 到位是,我们首先 |
| 106 | of all have Quad do a first pass review. | 四方首先通过审查 |
| 107 | What's very important in there is making sure that it has a way to actually try the experience | 里面最重要的就是确保它有办法去尝试经验 |
| 108 | out similar to a real user. | 与真正的用户类似。 |
| 109 | So that can mean end-to-end tests, opening up the experience, checking it in a browser, | 所以这可以意味着端到端测试, 打开体验,检查它在一个浏览器, |
| 110 | running any tools, doing any linting, doing any other checks that are necessary to build | 运行任何工具, 做任何衬垫, 做任何其他必要的检查以构建 |
| 111 | up that confidence. | 增强信心 提高信心 |
| 112 | People usually ask us, okay, well, what happens if you discover a bug or a gap in functionality? | 人们通常问我们, 好吧,好吧,如果你发现一个错误或功能上的漏洞,会发生什么? |
| 113 | Well, we have hooks that will then go and run on any edits. | 我们有钩子,然后在任何编辑上运行。 |
| 114 | They will study the diffs and allow us to get this loop in place where we are correcting | 他们将研究diffs 并让我们得到这个循环 在我们正在纠正 |
| 115 | the PR as quickly as possible. | 公关要尽快 |
| 116 | So we have Cloud Code doing the review. | 所以,我们有云代码 做审查。 |
| 117 | We have a code review product that you can also check out that's pretty good. | 我们有一个代码审查产品 你可以检查出来 这是相当不错的。 |
| 118 | And the idea there is we're trying to automate code review as much as possible. | 想法是,我们正在尝试 自动化代码审查 尽可能多的。 |
| 119 | This is something that we use ourselves. | 这是我们利用的 |
| 120 | And increasingly what's important is for your teams to be thinking about how can you cut... | 越来越重要的是 你们的队伍要思考如何切割... |
| 121 | We talk about taste and judgment so much these days, right? | 我们谈论的品味和判断这么多 这些天,对不对? |
| 122 | It's important for you to be trying to codify what you consider best practice and taste | 重要的是你必须努力 编纂你认为最佳做法和品味 |
| 123 | in your team's shared, like, quad.md, agents.md, anywhere where the agent can then consult | 在你团队的共享, 像,四分卫,代理,代理Md, 任何地方的代理然后可以咨询 |
| 124 | during code review. | 在代码审查期间。 |
| 125 | And then people are responsible for checking on the agent's evidence. | 然后人们负责检查特工的证据. |
| 126 | So how do we know? | 我们怎么知道? |
| 127 | If I'm building out a piece of UI that's going to be used by a user in the browser, I want | 如果我在浏览器里造出一个用户界面 供用户使用 |
| 128 | to make sure that it's tested in the browser, maybe generated me a video of that experience | 为了保证它在浏览器中测试, 也许生成了一个视频给我 体验 |
| 129 | where it's been clicking around and gives me confidence, well, okay, well, this works. | 它一直点击 并给我信心, 好了,好了,好了,这个工作。 |
| 130 | If I care about the mobile or desktop experience, maybe it generates me screenshots or videos | 如果我在乎移动或桌面体验,也许它会产生我截图或视频 |
| 131 | of that. | 就是这样 |
| 132 | But basically you want to get to a place where the agent's giving you as much verification | 但基本上,你想得到一个地方 在那里的代理给你 多少验证 |
| 133 | evidence as possible. | 尽可能提供证据。 |
| 134 | Cool. | 凉爽。 |
| 135 | Yeah. | 对 |
| 136 | Verification is key. | 核查是关键。 |
| 137 | I think you need to verify everything that the software factories will produce. | 我认为你需要核实软件厂生产的一切 |
| 138 | And that's what we are going to talk about today. | 这就是我们今天要谈的。 |
| 139 | So these stats, by the way, are from the state of code survey report that will be coming | 顺便说一句 这些数据来自即将到来的代码调查报告 |
| 140 | out very soon from Sonar. | 很快从声纳站出来 |
| 141 | We surveyed about more than 1,000 developers. | 我们调查了1000多个开发商。 |
| 142 | And we found some interesting data which will give us insights as to what's happening in | 我们发现了一些有趣的数据 将让我们深入了解 正在发生什么 |
| 143 | the industry today and where we are headed. | 今天的产业和我们的方向。 |
| 144 | So 84% say AI features gets features and fixes to production faster. | 因此,84%的人说AI的特性获得特性和修复以更快地生产. |
| 145 | The pull requests are getting larger and more complex. | 牵引请求变得越来越复杂. |
| 146 | And yet only 31% more code is reaching the production without review. | 然而,只有31%的代码没有经过审查就达到了生产。 |
| 147 | So that's a little bit of a scary fact, particularly for production applications in the regulated | 所以这有点吓人 特别是对于受管制的生产应用 |
| 148 | industries, in the mission critical applications, where you absolutely need to make sure that | 在任务关键应用中,你绝对需要确保 |
| 149 | you have to verify that code and you're confident in the code that is being shipped. | 你必须核实那个代码 你对正在运送的代码有信心 |
| 150 | So where are the next gains going to come from, Adi? | 阿迪,接下来的收益会从何而来? |
| 151 | I think the next gains are going to come from a combination of us really nailing verification, | 我认为接下来的成果将来自 我们的组合 真正钉钉核查, |
| 152 | as well as model quality. | 以及模型质量。 |
| 153 | But models are already, like if you take a look at Opus 5.5, model quality has already | 但模型已经,就像你看看Opus 5.5,模型的质量已经 |
| 154 | gotten to a place where it's pretty good. | 到了一个不错的地方 |
| 155 | And I would say that us being able to share best practices about that verification step | 我要说,我们能够分享核查步骤的最佳做法。 |
| 156 | is going to become increasingly more important. | 将变得越来越重要。 |
| 157 | Because if you tell people, yeah, just verify the work, their next question is, what does | 因为如果你告诉人们,是的,只是核实 工作,他们的下一个问题是,什么是 |
| 158 | that even mean? | 这甚至意味着? |
| 159 | And so I think that the conversation needs to evolve into specifically what kinds of | 所以我认为 谈话需要演变成 具体什么类型的 |
| 160 | We lean a lot on end-to-end tests, property tests, any visual verification tests, things | 我们非常依赖端到端的测试 财产测试 任何视觉验证测试 |
| 161 | that can help us just get that confidence to merge those changes in automatically. | 这能帮助我们获得信心 把这些变化自动合并 |
| 162 | Cool. | 凉爽。 |
| 163 | You mentioned light factories want to go dark, and that's the direction we are headed. | 你提到轻工厂想变黑 这就是我们的方向 |
| 164 | Our survey results show that today 5% of the companies are using software factories. | 我们的调查结果显示,今天5%的公司正在使用软件工厂。 |
| 165 | It will get to 34%. | 它会得到34%。 |
| 166 | And surprisingly, 39% in the financial services. | 令人惊讶的是,金融服务部门占39%。 |
| 167 | How many of you are using software factories today? | 今天有多少人在使用软件工厂? |
| 168 | Just here, yes, a handful of you. | 就在这里 是的 你们几个 |
| 169 | And how many are planning to use software factories by next year? | 有多少人计划明年使用软件工厂? |
| 170 | Yeah, that's a good question. | 是啊,这是一个很好的问题。 |
| 171 | How do you define a software factory? | 如何定义软件厂? |
| 172 | Would you like to take that question? | 你想接受这个问题吗? |
| 173 | Yeah, yeah, sure. | 是啊,是啊,当然。 |
| 174 | So we talked about definition of software factories just a little bit earlier. | 我们刚刚谈到软件厂的定义。 |
| 175 | A software factory is basically loops at scale. | 一个软件厂基本上是规模的循环. |
| 176 | And when we think about this, it's all about what end of the spectrum are you at in terms | 当我们想到这一点, 它的所有关于什么结束 光谱,你是在说 |
| 177 | of how much people are still involved in the process? | 有多少人仍在参与这一进程? |
| 178 | The more that the agents are able to generate verification and give you confidence, the | 特工人员越能进行核查并给予你信心, |
| 179 | more that you can get to that place where you go dark. | 更能到达黑暗的地方 |
| 180 | And I think that, yeah, most people are still somewhere in that spectrum and are not quite | 我认为,是的,大多数人 仍然在某个地方 在那种光谱,不完全 |
| 181 | as this data shows at that dark point just yet. | 正如这些数据显示的 在那个黑暗点刚刚。 |
| 182 | Yep. | 对 |
| 183 | Just to add to what Adi said, I think software factory, think of it like a production line | 为了补充阿迪的话 我认为是软件厂 把它当成生产线 |
| 184 | where you start defining the intent and your specifications. | 开始定义意图和规格 |
| 185 | Then you start writing the code. | 然后你开始写代码。 |
| 186 | So it's an assembly line where each of these software development lifecycle steps happen. | 因此,它是一个编组线,每个软件开发生命周期步骤都在此发生. |
| 187 | And automating them at scale, where there's pretty much in a dark factory, there would | 在黑暗的工厂里 实现它们的自动化 |
| 188 | be no lights on. | 没有灯光。 |
| 189 | And it's autonomously working, fully autonomously working. | 而其自主运作,完全自主运作. |
| 190 | A light factory is where people are involved in the process at every step or most steps. | 一个轻工厂就是人们在每一个步骤或大多数步骤都参与这一过程的地方. |
| 191 | Whereas a dark factory, the goal would be to turn off the lights and let the factory | 虽然一个黑暗的工厂, 目标将是关闭灯光 让工厂 |
| 192 | run independently 24 by 7. | 独立运行 24 by 7. |
| 193 | All right. | 好吧,我们走。 |
| 194 | So you talked about verification. | 所以,你谈到核查。 |
| 195 | I'll talk a little bit about what we think at Sonar about verification. | 我想谈谈我们在声纳公司对核查的看法。 |
| 196 | If you saw our keynote earlier today from our CEO, he talked about verification should | 如果你今天早些时候看到我们的首席执行官的主旨, 他谈到核查应该 |
| 197 | be multilayered. | 改为多层。 |
| 198 | We absolutely, 100% believe that each company should adopt a zero trust, multilayered verification | 我们百分之百相信,每个公司应该采取 零信任,多层次的核查 |
| 199 | approach. | 办法。 |
| 200 | And verification should be done in every loop, whether you're operating inside the agentic | 并且每个循环都要进行核查 无论你是否在特工内部运作 |
| 201 | loop or whether you're working inside the CI verification loop or whether you're doing | 循环,或者你是否在CI验证循环中工作,或者你是否这样做 |
| 202 | code maintenance. | 代码维护。 |
| 203 | In each and every loop, it is important to verify the output of the factories or the | 在每一个循环中,必须核查工厂或工厂的产出。 |
| 204 | agents that code that the agents are producing. | 特工们正在制作密码 |
| 205 | So we look at it in terms of layers and think of it like, you know, this first layer, which | 因此,我们从层次来看待它, 并想它像,你知道,这个第一层, |
| 206 | is a deterministic layer, which catches bugs, security issues, things like SQL injection, | 是一个决定性的层, 它捕捉bugs, 安全问题, 比如SQL注射, |
| 207 | buffer of bugs, and all of those are caught. | 虫子的缓冲器, 和所有这些被抓住。 |
| 208 | And then there's a second layer where agentic review happens, where agents are used to, | 还有第二层 代理审查发生 使用代理, |
| 209 | like Adi was talking about, agents review that code. | 就像Adi在说的那样 特工们审查那个代码 |
| 210 | And in that agentic review, they catch logic and intent issues, which the deterministic | 而在代理审查中,他们捕捉到逻辑和意图问题, 决定性的 |
| 211 | layer is not well suited to catch. | 层不适于捕捉。 |
| 212 | So we use each layer is targeted at catching a specific types of issues. | 因此,我们利用每一层都是为了抓住特定类型的问题。 |
| 213 | And then the third layer could be a software supply chain security layer, where you look | 然后第三层可能是软件供应链安全层,在那里你看起来 |
| 214 | at all the CVEs, all your malware detection happens, all your malicious package detection | 在所有的CVES, 所有你的恶意软件检测发生, 所有你的恶意包检测 |
| 215 | happens, all of that happens in the third layer. | 所有这一切都发生在第三层 |
| 216 | So it's a layered approach where each layer is tuned and optimized to catch a certain | 因此,它是一种分层的方法,每一层都经过调制和优化,以便捕捉到一定的 |
| 217 | type of issue. | 问题类型。 |
| 218 | So we absolutely recommend doing a multi-layered verification. | 因此,我们绝对建议进行多层次的核查。 |
| 219 | So there's another stat here, 28% companies have approved agents and shared controls and | 28%的公司已经批准了代理 和共享控制 |
| 220 | very few shared controls, and 5% are operating software factories autonomously. | 很少共享控制 5%的软件工厂在自主运作 |
| 221 | What are your thoughts there, Adi, in terms of? | 阿迪,你有什么想法吗? |
| 222 | So I've talked to so many companies where a couple of months ago, one of their biggest | 我和很多公司谈过了 几个月前他们最大的公司之一 |
| 223 | worries was they were allowing their engineering teams to go off and play around with agents. | 担心的是 他们允许他们的工程团队 与特工们玩耍 |
| 224 | And they got to a place where there were hundreds, thousands of agents in some cases, and they | 他们来到了一个地方 那里有几百,数千名特工 在某些情况下,他们 |
| 225 | just didn't know what they were doing. | 只是不知道他们在做什么。 |
| 226 | What are these agents actually doing? | 这些特工到底在做什么? |
| 227 | They're touching all of our systems, accessing all kinds of data. | 他们正在接触我们所有的系统, 访问各种数据。 |
| 228 | Maybe they're deploying to production. | 也许他们正在部署生产。 |
| 229 | And so a lot of enterprises think about concerns like agent identity and telemetry and being | 所以很多企业都考虑 代理身份 遥测 和正在 |
| 230 | able to track what's actually happening. | 能够追踪到底发生了什么 |
| 231 | And I'd say that the three biggest things for a lot of companies are, how do we think | 我想说对很多公司来说 最大的三件事是 我们如何看待 |
| 232 | about control? | 关于控制? |
| 233 | How do you think about permissions? | 你怎么看 许可? |
| 234 | And how do we think about the sandbox? | 沙盒怎么样? |
| 235 | How are we making sure that the agent isn't jumping out of the sandbox and doing things | 我们如何确保经纪人不会跳出沙盒 做事情 |
| 236 | that it shouldn't be doing? | 它不应该这样做? |
| 237 | Yeah. | 对 |
| 238 | At Sonar, we call it guide rails. | 在声纳,我们称之为导轨. |
| 239 | So instead of guard rails, we're thinking of calling it as guide rails, where we guide | 所以,我们不是在看守栏杆, 而是在考虑把它称为导轨, 在那里我们引导 |
| 240 | the agents to follow the rail tracks, a specific path to take and prevent them from going out | 追踪铁轨的特务 一条可走和阻止他们出轨的具体道路 |
| 241 | of bounds or, so to say, going out of the sandbox, like Adi was saying. | 就像阿迪说的 从沙盒里出去 |
| 242 | So it's very, very important to have guide rails for an agent and govern the agents that | 因此,非常,非常重要的是,有导轨 为特工,并管理特工 |
| 243 | you're working with. | 你和他们一起工作 |
| 244 | So we're coming to a close now. | 所以我们马上就要关门了 |
| 245 | As my timer is telling me, we just have a few minutes left. | 正如我的计时器告诉我的 我们只剩几分钟时间了 |
| 246 | Let's get into a lightning round. | 让我们进入闪电回合。 |
| 247 | One sentence response for each of these questions. | 以上各答一句. |
| 248 | And you have seen the questions already on the slide. | 你们已经看到幻灯片上的问题了。 |
| 249 | Can these be long sentences? | 这些会是长句吗? |
| 250 | They can be as long as the time permits. | 只要时间允许,它们就可以存在。 |
| 251 | Perfect. | 不错 |
| 252 | Light or dark factory for most teams by end of 2027, yes or no? | 到2027年末大多数球队的轻或暗工厂,是还是不是? |
| 253 | I think we're going to see light factories with some dark corners. | 我想我们会看到轻工厂 有一些黑暗角落。 |
| 254 | I don't think we're going to get to everybody in this room using dark factories entirely, | 我不认为我们会得到 在这个房间里的每个人 完全使用黑暗工厂, |
| 255 | but it's going to be something close to that. | 但它会是接近的东西。 |
| 256 | I would guess Anthropic would do it this year or next year. | 我想今年或明年Anthropic会这么做 |
| 257 | All right. | 好吧,我们走。 |
| 258 | Second question. | 第二问. |
| 259 | Should agent-written code be labelled in the PR? | 公关中应该贴上代理写法的标签吗? |
| 260 | I personally think it should. | 我个人认为应该这样 |
| 261 | Historically, we have felt a great deal of ownership around our code, pride around the | 从历史上看,我们一直 感受到很多所有权围绕我们的守则, 自豪围绕 |
| 262 | fact that, hey, I wrote this thing. | 事实上,嘿,我写了这个东西。 |
| 263 | I'm putting my name on it. | 我将我的名字在上面。 |
| 264 | 在Anthropic,我们想的有点不同。 | |
| 265 | We've evolved to the point where we don't really mind so much if Claude is labelled | 我们进化到这样 如果Claude被贴上标签 我们真的不介意 |
| 266 | as, hey, I'm the one that wrote this thing, because we're a little bit more focused on | 因为,嘿,我是写这个东西的人,因为我们更专注于 |
| 267 | the outcomes and making sure that the code, the features, what we're shipping to users | 和确保代码,特征, 我们正在运送给用户 |
| 268 | is the right thing. | 这是正确的事情。 |
| 269 | And so I personally think it's fine for the code to be written, labelled as coming from | 所以我个人认为 写代码是好的,标签是来自 |
| 270 | an agent, and that can also help us with things like figuring out, OK, well, if the agent | 也能帮我们找出问题 |
| 271 | wrote it, what are some of the constraints and extra checks that we need to keep in place? | 写道,我们需要保持哪些限制和额外检查? |
| 272 | I have the same point of view, Adi. | 我也有同样的观点 阿迪 |
| 273 | I think in the short run, at least in the foreseeable six to nine months, we should | 我认为,从短期来看,至少在可预见的六至九个月里,我们应该 |
| 274 | label the PRs till we are confident enough in our factories. | 把公关贴上标签 直到我们有足够的信心 在我们的工厂。 |
| 275 | We do want to know and track it. | 我们确实想知道并跟踪它。 |
| 276 | And it's also important for the financial customers here in terms of auditability. | 这对于这里的金融客户在可审计性方面也很重要. |
| 277 | If you label it, you know who's writing the code, because we are, at the end of the day, | 如果你贴上标签,你知道是谁在写代码, 因为我们是,最后, |
| 278 | still accountable. | 仍然负责。 |
| 279 | Yeah, absolutely. | 当然 |
| 280 | Accountability is the most important part here. | 问责制是这里最重要的部分。 |
| 281 | Like, we still label, like, who is the owner for this change? | 像,我们仍然标签, 像,谁的主人是这个变化? |
| 282 | I might have kicked off a quad run to complete a piece of work, but I'm still on the hook. | 我本可以开四步跑来完成一个工作 但我还是被钩住 |
| 283 | If something goes wrong, I have to go and fix it. | 如果出了问题,我必须去修复它。 |
| 284 | Yep, yep, key point, yep. | 是啊,是啊,关键点,是啊。 |
| 285 | The next question, actually, I'll improvise it here, because this is a leadership session. | 下一个问题,实际上,我在这里即兴回答, 因为这是一个领导会议。 |
| 286 | One skill a junior developer or leader should build on this year? | 一个初级开发者或领导者应该在今年的基础上发展一种技能? |
| 287 | I think that you should push your teams, if you're a leader, you should push your teams | 我认为你应该推动你的团队, 如果你是一个领导,你应该推动你的团队 |
| 288 | to really think hard about what verification means. | 认真思考核查的意义。 |
| 289 | How are you already approaching things like testing? | 你怎么已经接近的东西 像测试? |
| 290 | What is the gap between that? | 这两者之间有什么差距? |
| 291 | And you getting to a place where you're very comfortable with the agents being able to | 你得到一个地方,在那里你非常舒适 特工能够 |
| 292 | sign off on the code, being good enough to merge on their own. | 在代码上签名 足够自己合并 |
| 293 | And so figuring that out, if you're a junior, it's the other side of that, like, actually | 所以想清楚,如果你是低年级, 那其实是另一边 |
| 294 | getting more experience in the different flavors of tests, whether it's, like, end-to-end tests, | 获得更多经验 不同的口味的测试, 无论它是一样,端到端的测试, |
| 295 | property tests, et cetera. | 财产检验等。 |
| 296 | Yep, and the last, you're actually not following the rules, you're going beyond one sentence, | 是的,最后,你实际上没有遵守规则, 你将超越一句话, |
| 297 | but that's cool. | 但是这很酷。 |
| 298 | We have 30 seconds left, I'm fine. | 我们还有30秒,我没事 |
| 299 | We can go a minute over, I'm sure. | 我们可以过去一分钟,我确定。 |
| 300 | The agent opened the PR, it is safe to merge when... | 特工打开了公关,当... |
| 301 | You have full confidence nothing's going to go wrong, or that you've mitigated that as | 你完全有信心 不会出任何差错 或者你已经减轻了 |
| 302 | much as possible. | 尽可能 |
| 303 | Love it. | 喜欢它。 |
| 304 | All right. | 好吧,我们走。 |
| 305 | Here's my last slide and then we'll wrap it up. | 这是我最后一张幻灯片 然后我们把它包起来 |
| 306 | So what separates teams that are shipping fast but not shipping blind? | 那么,是什么将那些快速运输但并非盲目的运输的队伍分开? |
| 307 | I think that three key takeaways for you. | 我认为这三件关键外卖给你。 |
| 308 | Writing code is the fastest part and that is solved. | 写入代码是最快的部分,它得到解决。 |
| 309 | Strong review and verification for all teams is what will help you maintain that speed. | 对所有小组进行严格的审查和核查,将有助于保持这一速度。 |
| 310 | Fast to write is not fast to ship. | 快写不是快船 |
| 311 | You have to basically count what is getting merged, look at what is getting merged and | 你基本上要计算什么是合并, 看看什么是合并和 |
| 312 | make sure it doesn't escape the sandbox and the things like that before it gets into the | 确保它不会逃脱 沙盒和诸如此类的东西 进入之前 |
| 313 | production. | 生产。 |
| 314 | Verify everything. | 核实一切. |
| 315 | And autonomy is earned, we will get to software factories very soon, if not next year. | 有了自主权,我们很快就会到达软件厂,如果不是明年。 |
| 316 | You need to put in independent automated verification inside each loop of your factory. | 你需要在工厂的每一圈内 进行独立的自动核查 |
| 317 | Make sure that you can trust the output that the factory is producing. | 确保您可以信任工厂生产的产品。 |
| 318 | With that, I'll wrap it up. | 有了它,我会把它包起来。 |
| 319 | Thanks for attending our talk today, a conversation with Adi. | 谢谢你参加今天的谈话 和阿迪的谈话 |
| 320 | Again, Adi, appreciate you taking the time and joining us here in this conversation. | 再说一遍 阿迪 谢谢你抽空和我们一起谈话 |
| 321 | Thanks for having me. | 谢谢你邀请我 |
| 322 | Visit us at the Big Red Booth, the Sonar Booth in the Expo Hall if you have any further questions | 参观大红布斯 博览会大厅的声纳布斯 如果你还有问题的话 |
| 323 | or would like to talk about it. | 或者想谈谈 |
| 324 | Thank you again. | 再次感谢你 |
该视频共有字幕 324 条。解锁更多字幕为会员功能，请移动到 价格