在Notion，Codex正在改变工程师构建产品的方式。公司正在重新思考其构建的软件基元和抽象层，以便智能体能够使用它们。在招募新工程师加入团队时，他们看重的是好奇心和开放思维，因为该领域通常要求的多年经验目前尚不存在。而那些多年未编写生产代码的管理者，如今也重新回到代码库，与团队一起交付产品。
Ryan Nystrom负责Notion的AI产品工程。在过去一年多的时间里，他的团队构建或涉及了产品中几乎所有AI功能。他们发现，借助Codex，他们能够承担的任务范围解锁了新的可能性。
“我欣赏Codex的一点是，它在实际构建之前会花时间弄清楚问题。结果是，它构建的代码通常能直接符合我们代码库的标准，而不需要我回头清理一大堆它的工作。”
——Ryan Nystrom，Notion AI产品工程
在网页端构建Notion的AI语音输入
Codex自主执行的能力在一个近期项目中得到了清晰体现。Ryan完全使用Codex将Notion的AI语音输入功能带到了网页端。
“当我们说话时，我们可以提供更多的上下文，”他说，“如果我在打字，我会考虑我的措辞、用词。通过向Notion用户提供这一功能，他们能够提出更自然的问题，并包含更多上下文。我们想把这个功能带到Notion AI中。”
当这个功能需求提出时，Notion的移动应用已经有了一个可用的版本，但桌面端和网页客户端还没有。尽管Ryan并不完全清楚该功能在移动端是如何工作的，但他还是把问题交给了Codex。他将其指向移动端代码库，清晰描述了它在网页端需要呈现的样子，并提供了一种验证结果的方法。Codex一次性返回了完整的网页端实现初稿，其代码风格与Notion的代码库规范高度一致，第二天就可以交付。
“如果两年前让我构建Notion的语音输入功能，这个项目可能需要我和另一位工程师花两周时间，”Ryan说，“有了Codex，我完全靠自己，大概三四个小时就完成了。”
“它花了很多时间探索我们的移动端代码，然后最终一次性返回并编写了整个功能。我第二天就交付了，并立即让用户开始测试。”
——Ryan Nystrom，Notion AI产品工程
工作方式的变化
Notion的工程师现在可以常规性地将一组任务和检查工作的方法交给Codex，然后在它运行时离开。“我发现自己花了很多时间编写这些规范文档，然后交给Codex让它去工作，”Ryan说，“老实说，我现在几乎不再手动写代码了。”
在Codex之前，团队中的每位工程师一次只能专注于一个任务，还要在会议和同事支持之间挤时间。现在，他们可以并行处理多个任务，在启动工作的同时，不会失去那些曾经是瓶颈的团队支持。
“我管理着一个团队，传统上管理者没有时间写代码，”他说，“我能在支持团队的同时独自构建一个功能，这太不可思议了。我已经管理团队五年多了，从未能如此深入地处理编码问题。”
现在，他只需排好一个任务，去参加一连串会议，回来时功能就已经完成了。或者，他和他的团队拥有一位能和他们一起交付产品的管理者，而Ryan则重新获得了那份他以为已经放弃的工作内容。这种可用性甚至延伸到了工作时间之外。他会在睡前提出一个研究问题，让Codex运行一整夜，醒来时就能收到一份完成的报告。
Codex重新定义了一个小团队（甚至是一个人团队！）能够交付的基线，并为Notion的工程师提供了更多空间，让他们专注于自己最关心的工作。
“每当我需要研究一个任务、修复一个bug或做一点小调整时，Codex就在那里，随时待命。基本上，我在Notion有一个24/7全天候的实习生。”
——Ryan Nystrom，Notion AI产品工程
At Notion, Codex is changing how engineers build. The company is rethinking the software primitives and abstractions it builds so that agents can use them. When bringing a new engineer onto the team, they're hiring for curiosity and open-mindedness, since the years of experience the field would normally call for don't exist yet. And managers who hadn't written production code in years are back in the codebase, shipping alongside their teams.
Ryan Nystrom runs AI Product Engineering at Notion. In his year-plus tenure, his team has built or touched nearly every AI feature in the product. They’re finding that turning to Codex has unlocked new possibilities for what they can take on.
“What I appreciate about Codex is that it takes its time to figure things out before actually building. The result is that usually what it builds is to our codebase’s standards off the bat, rather than me having to go back and clean up a bunch of its work.”
—Ryan Nystrom, AI Product Engineering, Notion
Building Notion’s AI voice input on the web
Codex’s ability to autonomously execute showed up clearly in a recent project. Ryan exclusively used Codex to bring Notion’s AI voice input to the web.
"When we talk, we can provide so much more context," he says. "If I'm typing, I'm thinking about my prose, what words I'm using. By giving this feature to users on Notion, they're able to ask more organic questions and include a lot more context. We wanted to bring that to Notion AI."
When the feature request came in, Notion's mobile app already had a functional version, but the desktop and web client didn't. Even though Ryan wasn't entirely sure how the feature worked on mobile, he was able to give the problem to Codex. He pointed it at the mobile codebase, gave it a clear description of how it would need to look on the web, and provided a way to verify the result. Codex came back with a complete first cut of the web implementation, in one shot, that matched Notion's codebase conventions closely enough to ship the next day.
"If I were to build the Notion voice input feature two years ago, this is a project that would've taken me and maybe another engineer two weeks," Ryan says. "With Codex, I was able to build this in maybe three or four hours, entirely by myself."
“It spent a bunch of time exploring our mobile code, and then finally came back and wrote the entire feature basically in one shot. I shipped it the next day and immediately started letting users test it.”
—Ryan Nystrom, AI Product Engineering, Notion
How the work has changed
Engineers at Notion can routinely hand Codex a set of tasks and a way to check its work, then step away while it runs. "I've almost found myself spending a lot more time writing these spec documents that I can hand to Codex and let it work on," Ryan says. "Honestly, I don't really write code by hand anymore."
Before Codex, each engineer on the team could really focus on only one task at a time, squeezed between meetings and supporting peers. Now they're running multiple tasks in parallel, firing off work without losing any of the team support that used to be the bottleneck.
"I manage a team of people, and traditionally managers haven't had time to write code," he says. "The fact that I can build a feature solo while still supporting my team is crazy. I've been managing for five-plus years and never been able to go this deep on coding problems."
Now he can simply queue up a task, head into a block of meetings, and come back to a finished feature. Or he and his team get a manager who can ship alongside them, and Ryan gets back a part of the job he'd assumed he had traded away. This availability runs past the workday, too. He'll pose a research question before bed, let Codex run overnight, and wake up to a finished report.
Codex has reset the baseline for what a small team (even a team of one!) can ship, and given Notion's engineers more room to focus on the work they care about most.
“Whenever I need to research a task, fix a bug, or make a little tweak, Codex is just there, ready and willing. Basically, I’ve got an intern available at Notion 24/7.”
—Ryan Nystrom, AI Product Engineering, Notion
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文