两周前，我写了一篇关于 Claude Code 如何席卷 AI 世界的评论，文中提到“到 2026 年底，软件工程的面貌将发生巨大变化”。那篇文章捕捉到了 Claude 作为工具和产品的强大之处，我至今仍坚持这一观点，但它低估了我们在与软件相关的职业中使用这些产品的方式即将发生的变革。
更个人化的视角是：“如果我的工作符合 Claude 的形态，我更愿意用它来完成；很快我就会调整自己的方法，以便 Claude 能够提供帮助。”自那篇文章发表以来，我越来越强烈地感到，把我过去几年的工作方式直接套用到与智能体协作上，从根本上就是错的。在智能体时代，今天的习惯会让我因过度微观管理而限制自身提升，让自己疲惫不堪，并且把智能体局限在过于琐碎的任务上。更好的做法应该是更开放、更宏大、更异步。
我还不清楚该给自己开什么“药方”，但我知道前进的方向，也知道探索是我的职责。这个方向似乎涉及减少工作量，花更多时间培养内心的平静，这样大脑才能发挥最佳的指挥作用——让智能体去完成大部分艰苦的工作。
自从尝试使用搭载 Opus 4.5 的 Claude Code 以来，我的工作生活已逐渐转向适应与智能体协作的新方式。这种新的工作风格感觉比学习使用基于聊天的 AI 助手的时代转变更大。ChatGPT 让我能即时获取相关信息或我正在处理问题的潜在解决方案。而 Claude Code 则让我思考：既然我知道 AI 可以独立解决或实现许多子组件，那么我现在应该做什么工作？
每个工程师都需要学习如何设计系统。每个研究人员都需要学习如何运营实验室。智能体将人类推向了组织架构的上层。
我觉得自己因为较早赶上这波浪潮而占据优势，但不再认为仅仅靠努力就能维持长久的竞争力。当我可以让多个智能体在我的项目中并行高效工作时，我的角色正逐渐从使用电动工具转向指挥大军。更有效地指挥这些智能体，远比我多花几个小时埋头钻研一个问题要有用得多。
我目前默认的工作流程是：用 GPT 5 Pro 做规划，用 Claude Code 搭配 Opus 4.5 来执行。当 Claude Code 遇到难题时，我常常让它把信息传回给 GPT 5 Pro，配合非常详细的提示词进行深度搜索。仅靠 Codex 搭配 GPT 5.2 在极高思考强度下，感觉能力已经很强，甚至比 Claude 还要细致，但我还没找到发挥其最佳效果的方法。GPT Pro 本身感觉就像一个被困在错误用户界面中的强大智能体——它需要能够思考更长时间，并且需要一个专门处理研究任务的工作空间。
似乎我所有的朋友（包括那些名义上“非技术背景”的）都已经接受了这样一个事实：Claude 可以快速为你构建出令人惊叹的定制软件。Claude 将我过去的一个研究项目更新到了 uv，使其更易于维护；为我的 Discord 制作了一个验证机器人；为我的 RLHF 书籍绘制了大量图表；几乎要在我们的强化学习研究代码库中落地一个重要功能；还完成了无数其他本会耗费我数天时间的任务。这是当下最热门的事情——告诉你的朋友和家人你用 Claude 做了哪些小玩意儿。但这还远远低估了即将到来的东西。
我已经习惯让 Claude Code 实例在我的 DGX Spark 上持续运行，在我吃晚饭或上班时，尝试为我们的强化学习代码库实现新功能。它们会犯错，但能纠正自己大部分的错误，而且速度也相当慢，不过它们确实有能力。我迫不及待想回家看看我的那些 Claude 们又干了些什么。
Interconnects 是由读者支持的出版物。欢迎考虑成为订阅者。
我挥之不去的感受，是一种强烈的紧迫感：要把我的智能体从处理玩具级软件，转向执行有意义的长期任务。我们知道 Claude 能为我们完成数小时、数天甚至数周的有趣工作，但如何将这些砖块堆砌成连贯的长期项目？这是下一个工作时代的关键技能。
关于如何在前沿领域使用智能体，没有任何提示或指南——唯一的办法就是亲自上手。不要只让它们做清理工作，交给它们你最困难的任务之一，看看它会在哪里卡住，看看你能用它做什么。
软件正在变得免费，而在研究、设计和产品领域做出良好决策的能力，从未像现在这样有价值。
今天，善于使用 AI 比努力工作更能构成护城河。
以下是一些我认为恰当地探讨了即将到来的浪潮，或详细描述了使用智能体实际做法的文章。AI 领域这么多我尊敬的思考者同时聚焦于一个单一的新工具、一个过渡期和一种巨变感，这种情况实属罕见：
Import AI 441：我的智能体正在工作。你的呢？这篇文章促使我写下这些，并让我意识到这一刻有多么重要。
关于 AI 编程智能体的超高效能——重要的是，这篇文章写于 Claude Opus 4.5 之前，而后者是一次重大的阶跃变化。
Tim Dettmers 谈使用智能体：使用智能体，还是被抛在后面？
Steve Yegge 在 Latent Space 上谈氛围编程（以及如果你不理解如何做，你将如何被抛在后面）。
在智能体之间——为什么编程智能体不只是为程序员准备的。
1
这个可爱的 Clawd Bot 非常受欢迎，但我还没试过。
Two weeks ago, I wrote a review of how Claude Code is taking the AI world by storm, saying that “software engineering is going to look very different by the end of 2026." That article captured the power of Claude as a tool and a product, and I still stand by it, but it undersold the changes that are coming in how we use these products in careers that interface with software.
The more personal angle was how “I’d rather do my work if it fits the Claude form factor, and soon I’ll modify my approaches so that Claude will be able to help.” Since writing that, I’m stuck with a growing sense that taking my approach to work from the last few years and applying it to working with agents is fundamentally wrong. Today’s habits in the era of agents would limit the uplift I get by micromanaging them too much, tiring myself out, and setting the agents on too small of tasks. What would be better is more open ended, more ambitious, more asynchronous.
I don’t yet know what to prescribe myself, but I know the direction to go, and I know that searching is my job. It seems like the direction will involve working less, spending more time cultivating peace, so the brain can do its best directing — let the agents do most of the hard work.
Since trying Claude Code with Opus 4.5, my work life has shifted closer to trying to adapt to a new way of working with agents. This new style of work feels like a larger shift than the era of learning to work with chat-based AI assistants. ChatGPT let me instantly get relevant information or a potential solution to the problems I was already working on. Claude Code has me considering what should I work on now that I know I can have AI independently solve or implement many sub-components.
Every engineer needs to learn how to design systems. Every researcher needs to learn how to run a lab. Agents push the humans up the org chart.
I feel like I have an advantage by being early to this wave, but no longer feel like just working hard will be an lasting edge. When I can have multiple agents working productively in parallel on my projects, my role is shifting more to pointing the army rather than using the power-tool. Pointing the agents more effectively is far more useful than me spending a few more hours grinding on a problem.
My default workflow now is GPT 5 Pro for planning, Claude Code with Opus 4.5 for implementation. I often have Claude Code pass information back to GPT 5 Pro for a deep search when stuck with a very detailed prompt. Codex with GPT 5.2 on xhigh thinking effort alone feels very capable, more meticulous than Claude even, but I haven’t yet figured out how to get the best out of it. GPT Pro feels itself to be a strong agent trapped in the wrong UX — it needs to be able to think longer and have a place to work on research tasks.1
It seems like all of my friends (including the nominally “non-technical” ones) have accepted that Claude can rapidly build incredible, bespoke software for you. Claude updated one of my old research projects to uv so it’s easier to maintain, made a verification bot for my Discord, crafted numerous figures for my RLHF book, feels close to landing a substantial feature in our RL research codebase, and did countless other tasks that would’ve taken me days. It’s the thing de jour — tell your friends and family what trinket you built with Claude. It undersells what’s coming.
I’ve taken to leaving Claude Code instances running on my DGX Spark trying to implement new features in our RL codebase when I’m at dinner or work. They make mistakes, they catch most of their own mistakes, and they’re fairly slow too, but they’re capable. I can’t wait to go home and check on what my Claudes were up to.
Interconnects is a reader-supported publication. Consider becoming a subscriber.
The feeling that I can’t shake is a deep urgency to move my agents from working on toy software to doing meaningful long-term tasks. We know Claude can do hours, days, or weeks, of fun work for us, but how do we stack these bricks into coherent long-term projects? This is the crucial skill for the next era of work.
There are no hints or guides on working with agents at the frontier — the only way is to play with them. Instead of using them for cleanup, give them one of your hardest tasks and see what it gets stuck on, see what you can use it for.
Software is becoming free, good decision making in research, design, and product has never been so valuable.
Being good at using AI today is a better moat than working hard.
Here are a collection of pieces that I feel like suitably grapple with the coming wave or detail real practices for using agents. It’s rare that so many of the thinkers in the AI space that I respect are all fixated on a single new tool, a transition period, and a feeling of immense change:
Import AI 441: My agents are working. Are yours? This helped me motivate to write this and focus on how important of a moment this is.
on Hyperproductivity with AI coding agents — importantly written before Claude Opus 4.5, which was a major step change.
Tim Dettmers on working with agents: Use Agents or Be Left Behind?
Steve Yegge on Latent Space on vibe coding (and how you’ll be left behind if you don’t understand how to do it).
: Among the Agents — why coding agents aren’t just for programmers.
1
This cute Clawd Bot is very popular, but I haven’t given it a go yet.