A few weeks ago I looked at my screen time and did a double-take.

Last December, I was spending 13–16 hours a day staring at screens. Most of that was writing code. By March, I was down to 5–6 hours. And here’s the part that got me: my output in the first three months of 2026 was triple my entire 2025. Not triple the first quarter. Triple the full year.

Something structural had changed. I want to explain what — and I want to do it in a way that’s useful even if you’ve never written a line of code in your life. Because the shift wasn’t about code. It was about how I think about AI.

I. What Most People Actually Do With AI

If you use AI today, there’s a good chance your workflow looks something like this: you open ChatGPT (or Claude, or Gemini) in a browser tab. You type a question. You get an answer. You copy it somewhere. Maybe you paste an error back in and ask a follow-up. Then you close the tab and move on with your day.

This is how most people use AI. And while it works — Harvard/BCG research shows AI-assisted work on suitable tasks boosts speed by 25% and quality by 40% — it leaves most of the value on the table.

My friend YaGe calls this consumptive use. Each conversation starts from zero. The AI doesn’t know who you are, what you’re working on, or what you told it last Tuesday. You spend effort, but nothing accumulates. Nothing compounds.

The alternative is what he calls investment use: building a relationship with AI where every session makes the next one better. The AI remembers your context, your preferences, your goals — and it gets better at helping you over time because it knows more about your work.

The difference isn’t about which chatbot you use. It’s about whether you’re chatting or building.

II. The Thinking Shift

Here’s the question that changed everything for me.

When I hit a problem, I used to ask: How do I solve this?

Now I ask: What would an agent need to know to handle this for me?

That one question changes everything. Instead of asking AI a question and copying the answer, I’m designing a system. What context does it need? What should it have access to? What should it do on its own, and what should it bring to me for review?

Here’s the thing: this isn’t a technical skill. It’s a management skill. And that turns out to be the entire secret.

Andrej Karpathy — co-founder of OpenAI — captured this shift on the No Priors podcast. He said the new mode of working with AI isn’t coding. It’s “manifesting”: expressing your intent clearly and letting agents execute. You go from being the producer to being the director. Separately, he wrote on X: “It’s not magic, it’s delegation.”

If you’ve ever managed an employee, an intern, or even just explained a task to someone clearly enough that they could do it without you — you already have the core skill. The question is whether you’re applying it to AI.

The Manager Analogy (It’s Not a Metaphor)

Here’s something that keeps surprising people: management skills transfer to AI work better than pure engineering skills. Working with an AI agent isn’t like programming a computer. It’s like onboarding a new hire.

Think about what you do when a good new hire joins your team:

• You give them context — background on the project, the team, the goals
• You set boundaries — what they can decide on their own, what needs your approval
• You define success — not step-by-step instructions, but what “done” looks like
• You review their work — especially early on, before trust is established
• You build trust incrementally — small tasks first, bigger ones as they prove reliable

This is exactly how you should work with AI agents. The shift isn’t from “human does it” to “AI does it.” It’s from “I’m the one doing the work” to “I’m the one managing the work.”

What to Keep, What to Hand Off

Now, you shouldn’t delegate everything. Anthropic’s internal research found a practical rule of thumb their own engineers use:

1. Is the result easy to check? If you can glance at the output and tell whether it’s right — hand it off.
2. Is it repetitive? If you do the same thing every week — hand it off.
3. Is it low-stakes? If a mistake is easily fixed — hand it off.

Tasks that need taste, judgment, relationships, or high-stakes decisions? Those stay with you. One Anthropic engineer nailed it: “The more excited I am to do the task, the more likely I am to not use Claude.”

Delegate the work you dread. Keep the work that energizes you. That’s the whole framework.

III. From Chatbot to System: What Changes

Let me give you two concrete examples from my own life. Neither involves coding.

Example 1: Personal Investment Advisor

In chat mode, I activate the AI when I remember to. In system mode, the AI activates me when something needs attention. Push, not pull. I review. It runs.

Example 2: Personal Health Assistant

Notice the pattern. In both cases, the AI isn’t smarter. It has better context. It knows my situation, my goals, my history. And it doesn’t forget between sessions.

Don’t get me wrong — full “system mode” for finance and health is still early. Today’s tools get you partway there — persistent context, document uploads, scheduled check-ins — but continuous monitoring and proactive alerts are probably 12–24 months out for most consumer platforms. The direction is clear though. The infrastructure is catching up.

That’s the shift. From asking questions to building systems. From one-shot interactions to relationships that accumulate context over time.

IV. The Electric Motor Lesson

There’s a historical parallel here that I love.

When factories began electrifying in the early 1900s, factory owners did the obvious thing: they ripped out the central steam engine and dropped in a big electric motor. Same belt-and-shaft system. Same factory floor. New power source.

It helped. A little.

The real productivity explosion came a generation later, when someone realized you could put a small motor at each workstation. No more belts. No more central shaft. Each machine independent. The entire factory floor had to be redesigned from scratch.

YaGe’s observation: “Currently, we often operate in the ‘main shaft’ phase of the AI era.”

The San Francisco Federal Reserve made the same point in February 2026. Chair Mary Daly wrote: “Innovative firms used imagination and creativity to start fresh and build a world shaped by electricity, rather than leverage electricity in a steam-powered world.”

If you’re pasting questions into a chatbot window, you’re putting an electric motor on a steam-powered factory. The power source is new. The workflow is old. BCG found that companies who merely “deploy” AI onto existing processes see marginal gains. Companies who “reshape” their workflows around AI deliver 3.6× higher total shareholder return compared to those who don’t.

Most people are still deploying. The opportunity is in reshaping.

V. What a “Long-Running Agent” Actually Is

You’ve probably heard the word “agent” tossed around a lot. Let me make it concrete.

A chatbot answers one question at a time. It forgets you between sessions. It can only see what you paste into the window. It waits for you to start every interaction.

A long-running agent is different in four ways:

The key word is persistence. A long-running agent doesn’t start from zero every time. It accumulates context, builds on prior work, and gets better at helping you the more you use it.

Six months ago, using one of these required developer tools. That’s changed a lot. Here’s your actual pathway, from easiest to most powerful.

VI. Your Pathway: From Chat to System

Level 1: Give Your AI Memory (15 minutes, tonight)

The single biggest upgrade you can make is moving from throwaway conversations to a persistent workspace.

Why does this matter so much? Vercel tested this rigorously: embedding a compressed knowledge index in a persistent context file moved AI agent success rates from 53% to 100% on their eval suite. Same model. Same tools. Just persistent context. Anthropic’s engineering team summed it up: “Intelligence is not the bottleneck. Context is.”

Level 2: Let Your AI Work Autonomously (30 minutes, this week)

Once you’ve got persistent context, the next step is letting the AI do multi-step work without you hovering over every prompt.

Level 3: Design Your Own Agent Workflow (1 hour, this month)

Now, this is where you apply the thinking shift. It’s not about learning new tools — it’s about redesigning a workflow you already have. Pick one thing you do repeatedly — a weekly report, a research process, a client onboarding workflow. Instead of asking “how can AI speed this up?” ask:

1. What context does the agent need? — Background docs, templates, past examples, rules
2. What actions should it take autonomously? — Research, draft, organize, analyze, remind
3. What should it bring to me for review? — Decisions, final drafts, anything high-stakes

Write this down. You just designed an agent workflow. Now build it using whichever Level 1 or Level 2 tool fits your situation.

This is the step most people skip. You’re encoding your judgment — your conventions, your preferences, your domain expertise — into a form the AI can use every session. That’s what kills the cold-start problem: each session begins where the last one left off, instead of from zero.

VII. The Trust Curve

One more thing. Anthropic’s research team found something I love: people build trust with AI the same way they built trust with Google Maps.

You start by using it for unfamiliar routes — low-stakes, easy to verify. Then you trust it for familiar ones. Eventually you rely on it for complex multi-stop navigation you would never have attempted manually.

AI delegation follows the same curve. Small tasks first. Check the output. Calibrate your trust. Expand the scope. The people who try to skip straight to full autonomy get burned (I learned this the hard way — I wrote about it here). The people who build trust incrementally? They’re the ones who unlock the real leverage.

Atlassian quantified this. Simple AI users — those who treat it as a search engine — save about 53 minutes per day. Strategic users — those who build persistent workflows and iterate — save 105 minutes per day. Double the return, same tools.

Now What

Here’s the honest version. This isn’t about replacing yourself with AI. It’s about changing the type of work you do. When I cut my screen time in half and tripled my output, it wasn’t because each hour became six times more productive. It was because I stopped doing work that could be delegated and started doing work that only I could do. Strategy. Judgment. Decisions. The stuff that actually matters.

Anthropic found that roughly 27% of AI-assisted work consists of tasks that would not have been done otherwise. The output multiplier doesn’t come from doing the same things faster. It comes from doing things you never would have attempted.

My friend YaGe calls this “The Invisible Price List”: we all carry an unconscious model of what is “worth doing.” AI tears up that price list. When the cost of research, analysis, and organization drops to near-zero, you can just... explore every path and then decide. The effort is disposable. The insights compound.

So here’s what I’d do this week, if I were you:

1. Tonight: Set up one persistent workspace (Claude Project, Custom GPT, or Workspace Studio). Upload your context. Write your instructions. Stop starting from zero.
2. This week: Pick one repeating workflow and design it as an agent task. Write down: what context, what actions, what comes back to me.
3. This month: Try letting the agent work autonomously on something real. Review the output. Calibrate your trust. Expand.

The tools are here. The thinking is what’s missing.

Transparency note: I put every claim in this article through a systematic devil’s advocate review. Three factual errors were corrected, one fabricated claim was removed, and several recommendations were revised. If you want to see what broke and what held up, the full report is there.

前几天我看了一眼screen time，吓了一跳。

去年12月，我每天盯屏幕的时间大概是13到16个小时。大部分时间花在写代码上——IntelliJ、VS Code、PyCharm，就是抠着IDE干活。到了今年三月，降到了5到6个小时。

但真正让我坐不住的是这个：2026年头三个月，我的产出是整个2025年的三倍。不是三倍于去年同期，是三倍于去年一整年。

时间砍了一半，产出翻了三倍。这不是效率提升——这是工作方式变了。我想把这个变化讲清楚，而且讲成哪怕你一行代码都没写过也能用的东西。因为这个转变的核心其实不是代码，是思维方式。

一、大多数人其实在「浪费」AI

你现在用AI的方式，大概率是这样的：打开浏览器，开一个ChatGPT或者Claude的标签页，输入一个问题，拿到回答，复制粘贴到别的地方。可能再把报错贴回去问一下。然后关掉标签页，继续干别的。

这是绝大多数人用AI的方式。这个方式有用吗？有——哈佛/BCG的研究显示在适合AI的任务上，速度能提升25%，质量提升40%。但这只是冰山一角。

我朋友YaGe管这叫消耗型使用。每次对话从零开始，AI不知道你是谁，不知道你在做什么，不知道你上周二跟它说过啥。你花了时间精力，但什么都没积累下来。

反过来呢，是投资型使用：你跟AI之间是一个持续的关系，每次对话都让下一次更好。AI记住你的背景、你的偏好、你的目标。用得越多，它帮你越准。

区别不在于你用的是哪个chatbot。区别在于你是在聊天，还是在搭系统。

二、思维方式的转变

有一个问题，彻底改变了我的工作方式。

以前遇到事情，我的第一反应是：我怎么解决这个？

现在我想的是：一个agent需要知道什么，才能替我处理这件事？

就这一个问题，把我跟AI打交道的方式全变了。我不再是问AI一个问题然后复制答案——我是在设计一个系统——它需要什么context？需要access哪些信息？哪些事它可以自主做，哪些需要拿回来让我review？

这不是一个技术能力，这是一个管理能力。说白了，这就是全部的秘密。

Andrej Karpathy——OpenAI的联合创始人——在No Priors播客上说了个很准的东西。他说现在跟AI工作的方式不是coding，是“manifest”：把你的意图清晰地表达出来，让agent去执行。人的角色从「干活的人」变成了「导演」。他在X上另一个帖子里写得更直接：“It’s not magic, it’s delegation.”

如果你带过团队、管过人、甚至哪怕只是把一件事讲清楚让别人能不问你就干完——你就已经有了核心能力。问题只是你有没有把这个能力用到AI上。

管理这个类比，其实不是类比

有一个事挺反直觉的：实际操作中，管理能力用在AI上往往比纯工程能力好使。跟AI agent合作，不像在编程，更像在带新人。你想想看，你带一个新人入职会干什么：

• 给背景——项目是干嘛的，团队是什么情况，目标是什么
• 定边界——什么事他可以自己决定，什么事必须来问你
• 定义完成标准——不是step by step的指令，是「做成什么样才算做好了」
• Review产出——尤其是刚开始，信任还没建立的时候
• 逐步放权——先给小活，做好了再给大的

跟agent合作，一模一样。转变不是从「人干」到「AI干」，是从「我是干活的人」到「我是管活的人」。

什么留给自己，什么交给agent？

不是所有事都应该交出去。Anthropic自己的工程师总结了一个挺好使的判断框架：

1. 结果容不容易检查？你扫一眼就能判断对不对的——交出去。
2. 是不是重复性的？每周都要做的——交出去。
3. 风险高不高？错了容易改的——交出去。

需要品味、判断力、人际关系或者高风险决策的事？留给自己。Anthropic有个工程师说了一句挺到位的：“我越想亲自做的事，越不会用Claude。”

你不想干的活，交出去。你真正有热情的活，留着。就这么简单。

三、从问答到系统：到底变了什么？

我举两个自己的例子。都不涉及写代码。

例子1：个人投资顾问

聊天模式是我想起来才去问它。系统模式是它在需要的时候来找我。信息流从pull变成了push。我review，它执行。

例子2：个人健康助理

你注意到pattern了吗？两个例子里，AI都没变聪明。变的是它拿到的context——它知道我的情况、我的目标、我的历史。而且它不会在两次对话之间全忘了。

但说个实际的：金融和健康的「系统模式」目前还在早期。现在的工具能做到持久化context、上传文档、定期检查——但连续监控和主动提醒，大部分消费级平台还要12到24个月。方向是清楚的，基础设施在追赶。

这就是那个转变。从问问题，到搭系统。从用完就扔，到持续积累context。

四、电动机的教训

讲一个历史上的事，我觉得挺有意思。

1900年代初期电力开始进入工厂的时候，工厂老板做了一件最直觉的事：把中央蒸汽机拆了，换上一个大电动机。皮带传动系统不变，车间布局不变，就是换了个动力源。

有用吗？有一点。

但真正的生产力爆发是一代人以后的事。有人想明白了，可以给每个工位装一个小电机。不需要皮带了，不需要中央传动轴了，每台机器独立运转。整个车间布局要从头设计。

YaGe的判断是：“目前，我们大多数人还处在AI时代的‘主轴’阶段。”

旧金山联储2026年2月的经济通讯里，Mary Daly写了一句话：“真正的创新者是从零开始，围绕电力重新设计一个世界——而不是把电力塞进蒸汽动力的世界里。”

如果你现在的用法是在浏览器里粘问题给chatbot，那你就是在蒸汽工厂里装电动机。动力源是新的，工作流是旧的。BCG的研究发现，只是把AI「部署」到现有流程上的公司，收获很有限。而那些围绕AI能力「重塑」工作流的公司，股东总回报高出3.6倍。

大多数人还在「部署」。机会在「重塑」。

五、Long-Running Agent到底是什么？

你可能最近到处听到「agent」这个词。我把它讲具体。

一个chatbot，一次回答一个问题，下次对话不记得你，只能看到你粘进去的东西，需要你发起每一次交互。

一个long-running agent，跟它有四个本质区别：

核心就一个词：持续性。Long-running agent不是每次从零开始的。它积累context，在之前的工作基础上继续，用得越多越好用。

半年前，要用这种东西还需要developer tools。现在不需要了。下面是你实际可以走的路径，从最简单的到最强大的。

六、你的路径：从聊天到系统

Level 1：给你的AI装上记忆（15分钟，今晚就能做）

你能做的最大升级，就是从一次性对话换到一个持久化的workspace。

这为什么重要？Vercel做过严格测试：一份写得好的context文档，把AI agent的成功率从53%拉到了100%。同一个model，同样的工具，只是加了持久化的context。Anthropic的工程团队说了句特别到位的：“Intelligence is not the bottleneck. Context is.”

Level 2：让AI自主干活（30分钟，这周搞定）

有了持久化的context之后，下一步是让AI做多步骤的工作，不用你一句一句地喂。

Level 3：设计你自己的Agent Workflow（1小时，这个月内）

到这一步，前面说的思维转变才真正用上了。选一件你反复在做的事——周报、research流程、客户onboarding——然后不要问「AI怎么帮我加速这件事？」而是问：

1. Agent需要什么context？——背景文档、模板、历史案例、规则
2. 什么它可以自主做？——调研、起草、整理、分析、提醒
3. 什么需要拿回来给我看？——决策、终稿、任何高风险的东西

把这个写下来。写完你就会发现——你刚刚设计了一个agent workflow。接下来用Level 1或Level 2的工具去实现就行了。

这是大多数人跳过的那一步——「蒸馏」。你在把自己的判断力——你的习惯、你的偏好、你的领域专业知识——变成AI每次都能复用的东西。这一步，让每次session都从上一次的基础开始，而不是从零开始。

七、信任曲线

还有一点。Anthropic的研究团队发现，人信任AI的过程，跟当年信任Google Maps的过程一模一样。

你一开始只在不熟悉的路线上用——风险低，容易验证。然后在熟悉的路线上也开始用。最后你会靠它去做复杂的多点导航，那种你以前想都不会想的路线。

AI delegation也是这个曲线。先给小活。检查产出。校准你的信任。逐步扩大范围。直接跳到全自动的人会摔跟头（我自己就摔过——写在这里了）。逐步建立信任的人，才能真正leverage起来。

Atlassian有一组数据挺说明问题的。随便用用AI的人——当搜索引擎使——每天省53分钟。认真搭了持久化workflow、反复迭代的人——每天省105分钟。同样的工具，两倍的回报。

那接下来怎么做？

讲真。这不是要用AI把你自己替换掉。关键在于改变你干的活的种类。我把screen time砍了一半、产出翻了三倍，不是因为每小时变成了六倍效率。是因为我不再干那些可以被delegate的活了，开始干只有我能干的活。战略。判断。决策。真正重要的东西。

Anthropic的数据显示，大概有27%的AI辅助工作，是以前根本不会去做的事。产出翻倍不是因为同样的事做得更快了，是因为你开始做以前想都不会想的事。

YaGe管这叫「隐形价格表」：我们脑子里其实都有一个隐性的判断，觉得什么事「值得做」什么事「不值得」。AI把这个价格表撕了。当调研、分析、整理的成本降到接近零，理性的做法就是把每条路都探一遍，然后再做决定。反正精力花了就花了，但洞察是能积累下来的。

所以如果我是你，这周这么做：

1. 今晚：建一个持久化的workspace（Claude Project、Custom GPT、或Workspace Studio）。上传你的context。写好instructions。别再从零开始了。
2. 这周：选一个重复性的workflow，按agent任务的思路设计它。写下来：什么context，什么动作，什么拿回来给我。
3. 这个月：让agent真正自主地做一件实事。Review产出。校准信任。扩大范围。

工具已经在了。大多数人缺的是思维方式。

透明说明：这篇文章的每一个数据和建议都经过了系统性的魔鬼代言人审查。三个事实错误被修正，一个AI起草时编造的说法被删除，多项建议被修订。想看哪些站住了、哪些没站住，完整报告在那里。