# Agentic Engineering Setup (after 2,000+ hours)
**作者**: David Ondrej
**日期**: 2026-08-31T13:59:59.000Z
**来源**: [https://x.com/DavidOndrej1/status/2094424967345496191](https://x.com/DavidOndrej1/status/2094424967345496191)
---
Over the past 3 years, I've spent well over 2,000 hours coding with AI, and I've personally interviewed some of the most productive people in the space of Agentic Engineering.
Below is my full Agentic Engineering setup as it currently stands, Q3 of 2026.
## Interface
This means the UI / CLI how you interact with agents. My main interface is bb.
It's open source, completely free, and it allows you to use any subscription, any agent, any model inside of a single GUI. Codex, Claude Code, Pi, Cursor CLI, OpenCode, Grok Build, Hermes, all within the same UI.
The problem with apps like Codex or Cursor is that they only allow their own models and their own subscription. The goal is getting the most tokens for the least amount of dollars.
All the features you like from the Codex or Cursor apps are inside of bb, and it's improving every single week (plus, it's fully open-source and 100% free to use).
Another thing I use a lot is cmux.
When you launch a new cmux workspace, you can divide the screen just like you would in tmux (that's why the similar name), launch different terminals in every pane, and there's a built-in browser.
Where cmux breaks is once you have a lot of agents and workspaces. The left sidebar really is not the right primitive. Fine for a couple of things going on, but for serious agentic engineering work at scale it's not the best.
I use Ghostty as my terminal because it's very fast and native.
Inside of Ghostty, you can run Herdr, which is basically tmux but for agents, a backend runtime for agents. Very minimal, very lightweight, lives in the terminal, and when an agent is finished, its state shows on the left: done, idle, blocked, running.
Tracking the states of AI agents is ESSENTIAL. My prediction is that in 3 to 6 months this "agent state tracking" becomes more and more important, because you won't be talking to a single agent. You'll be talking to a manager agent that manages a lot of worker agents.
The last interface I have to mention is Corral, something I developed myself.
Instead of randomly switching between agents whenever they finish (in Herdr, there's no real order), every agent has a priority. Just like tasks have different priority/importance levels. When a P1 agent finishes, he goes to the top. You should never respond to a P4 agent when a P1 agent has finished running. (yes... i do need to make corral open-source. haven't got to it yet.)
## Models and Subscriptions
You want the most tokens for the least amount of dollars possible. This should be one the main goal of every Agentic Engineer (after getting shit done).
There are 4 main subscriptions right now, and yes... this could be completely different 2 months from now.
The best "bank for your buck" deal right now is OpenCode Go. It's only $10, and it gives you Kimi K3, Grok 4.6, GLM 5.3, DeepSeek V4 Pro, and many other models... But it doesn't have the best models like Fable 5 and GPT-5.6 Sol (plus, the usage limits are quite small)
So, if you have a bit more money, here's what you should do:
- $30 -- get OpenCode Go + ChatGPT Plus ($20)
- $50 -- Add the $20 Claude Code subscription on top.
- $70 -- Add Cursor's $20 a month, and you have the lowest tier of all subscriptions.
- $110 -- OpenCode + one of the big plans. ChatGPT $100 plan will give you a better deal than Claude. It is what it is. OpenAI has more compute, they're willing to subsidize it more.
- $210 -- Just get both $100 plans.
- And if you really are serious (like me) just get all $200 plans (Codex, Claude, Cursor) because those are 20x usage, and these plans give you the best deal.
btw... the Cursor plan is very underrated. Cursor, aka Grok, is going to become a great subscription because of the SpaceX acquisition. SpaceXAI has loads of compute, so they're allowed to play the game of subsidizing. And -- I think -- the Cursor/Grok plan gives you separate limits for Cursor + Grok Bot, which is just incredible.
Grok Bot is rapidly becoming the new way people interact with agents, so having a Cursor subscription has never been more important (not sponsored lol, it's just true). Plus, the new model -- Grok 4.7 -- is right around the corner.
No matter what, do not pay API pricing. It is the worst deal out there. Just get the subscriptions.
## Cloud Agents
It's quite obvious that cloud agents are the future. Cursor, Amp, Devin, Codex... all of these companies are betting everything on Cloud Agents.
The proof that cloud agents are the future is the graph below.
[image here]
This is Cursor's internal share of merged PRs from cloud agents: around 10-15% at the start of this year, approaching 60% now. And that's merged PRs, the stuff that actually gets used. Soon enough this will be 70%, then 80% and then 90%.
The issue with running all agents locally, on your machine, is that it's not scalable. You cannot run hundreds of agents at once. All it takes is a couple of agents deciding to run your entire test suite at the same time and your computer will begin to make weird noises (even my $7,000 macbook pro struggles).
Cloud Agents give you isolated environments, persistent sessions, durable access to internet and electricity. If you close your laptop, you lose your sessions. Lose internet for a couple of minutes, and the harnesses cannot self-recover.
The problem with the existing cloud agent solutions is the INSANE level of ecosystem lock-in. Setting up the environments and all your secrets is a lot of hours, and then you're locked in: your sessions are there, you're on their pricing, and you give them all your data. Even if they don't train models on it, there are so many other ways to use your data.
The solution is to have your own server. And thanks to AI, this takes like 10 minutes to set up (seriously). Just get a VPS, run Herdr on it, and SSH into it.
Herdr gives you persistent agent sessions, and SSH lets you connect from your phone, your laptop, anything. You can literally achieve the 80/20 of cloud agents for a couple of dollars, without the lock-in.
## Build your own Cloud env
I use Hostinger for my VPSs, and a KVM2 plan is enough. Here's the main thing I want to get across... you don't need to be an expert in VPSs, DevOps, Linux, none of that.
Just talk to your agent in plain English!!!
In the upcoming video, I set the whole thing up live. A cmux workspace, a coding agent in the left pane (Cursor CLI running Grok 4.6), an empty terminal pane on the right.
My cmux skill lets the agent find that other pane and run commands in it. I SSH'd into the fresh VPS myself, then told the agent "learn everything about that server, and set up the dev environment. Herdr, Node.js, Python 3, Git".
It analyzed the VPS in seconds, installed everything, started Herdr, then installed Pi Agent, found an OpenRouter key on my macbook, and went through the full setup itself. A few short prompts later, I had Pi running GPT-5.6 Sol and a second session running Fable, both in the cloud, on my own VPS, with full root access.
If something happens to my computer or wifi... If my MacBook explodes, those agents keep running. The full walkthrough is in the video. [Link to my YouTube here.](https://www.youtube.com/@DavidOndrej)
One more speed thing... I dictate with SuperWhisper. Most of you reading this probably type at 40 or 50 words per minute. That is very slow.
BUT! you can speak at 250+ WPM. A voice AI tool (like Superwhisper, Glaido, Whispr FLow) instantly makes you 3-4x faster at sending prompts. Use one. Don't be stupid.
## Harness
First harness I need to mention is Pi Agent, the GOAT.
The most minimal harness out there: just 4 tools, always runs in YOLO mode, supports any model, any provider. Very elegant, super configurable, and that's why so many people build on top of Pi. It's open source, completely free, just go to [pi.dev](//pi.dev) and get it. Non-negotiable. It's the first harness I put on the VPS.
Cursor CLI. Very underrated, because you can use all the models: Grok, GPT models, Anthropic models, Kimi. You can tag skills, and you can pre-send messages. Great harness overall.
The next category of Harnesses is what I like to call "self-improving" harnesses.
The two most popular ones are Hermes Agent and Prime Agent. This is for when you don't know what you're doing. If a task has a lot of uncertainty, a lot of figuring out to do, use a self-improving harness, because it creates skills and improves with you over time.
And finally, the classics, Claude Code and Codex. I have them as aliases. A lot of people type claude --dangerously-skip-permissions every single day. Extremely slow, extremely inefficient. I type cc and it launches Claude Code with permissions bypassed; cx launches Codex in YOLO mode.
YOU MUST create global aliases for the long commands you run often. That's one of the laws of agentic engineering: How can you get more done in the same amount of time?
## Skills
My skills repo went viral last month. (see [github.com/davidondrej/skills](//github.com/davidondrej/skills) ) It's also completely free, open source, all that.
The SKILLs most relevant to Agentic Engineering are:
*(1) /total-review
- It runs two other skills, /gpt-review and /fable-review, which review whatever code changes you just did with GPT-5.6 Sol and Fable 5, then dedupes both lists into a single list of the issues that actually matter. It's like asking all of your smartest friends to review your job application, and they only give you the biggest issues. Run it on medium-to-large changes, especially if a different model built it. If Grok 4.6 did the work, you want a totally different model reviewing it.
IMPORTANT: anything you repeat often enough should become a preset.
If it's a single step, use text replacements. I have them as Raycast snippets. "Answer in short in plain English." "Make your previous answer simpler and shorter." "Stage all files, write a clear commit, push to GitHub."
If it's a multi-step workflow, turn it into a skill.
(2) /ask-then-build
- I use this skill every single day, before any building. Instead of saying "make this Windows compatible" and letting the model silently make important architectural choices you might regret later, it walks you through the main decisions one at a time, with options. AI models are great at coding, great at implementation. They don't have taste. They don't have good judgment. You, as the human, need to stay in charge of that.
(3) /deepapi
- This skill is what I use for any deep research, any scraping, anything web. Codex and Claude Code come with basic web search, but no scraping, no deep research, and they get blocked easily. Run 8 fast web searches and give me the top 3 options, scrape Twitter, scrape GitHub, find 3 ways of contacting a person. Everyone on my team uses it. Must-have.
(4) Guardrails and push lock
- More boring, but absolutely essential, and you only set them up once. Global agent guardrails is a pre-tool-call hook that makes sure your agents never wipe your disk, never overwrite Git history, never touch your password manager. And push lock, for when you're running 15+ agents in parallel: one OS-level kernel lock on the whole ship. Merge, verify, push, CI, deploy, health check.
Don't install all of my skills. Just grab the ones you need.
## Worktrees
A worktree is basically a copy of your primary checkout into a separate folder and creates a new Git branch there, so agents can work in parallel, completely isolated.
On a small project, that's absolutely overkill. Stay on a single branch and work faster.
On medium-to-large projects where you're running 20-30+ agents at all times, there's no way to avoid it. Without worktrees, the agents will conflict, reverse each other's changes, and fight each other. Another good thing about BB: it has built-in worktrees. It remembers that on my big repos I always want a new worktree based off of origin/main.
## Other Agentic Engineering Tips
Know when to use each model.
- Designing a plan or starting a new project? Fable. It has the most sparks of genius. Fixing a deep, serious bug? GPT-5.6 Sol, max reasoning effort. Default chatting? Grok 4.6 on high. Almost the same intelligence, but 2x cheaper and 2x faster. Front-end? Kimi K3.
- And when a major new model is released, set aside a day where you only use that model. Don't listen to Twitter. Try it yourself.
Know when to review.
- I'm not running total review on every change. A small front-end tweak goes to prod right away.
- And NEVER do recursive reviews. If you tell a model "find the 5 biggest issues", it will find 5 issues even if the codebase is perfectly fine. These models invent imaginary bugs.
Pre-sending.
- Usually I know what the agent will do next, so I queue the messages ahead of time: "implement the plan", "run Fable review on this", "now fix those things".
- Sometimes it's 2 messages, sometimes 6.
- Never use a harness that doesn't allow you to pre-send.
Subagents are overused.
- A lot of people just burn their limits on them. I use them when I'm in control. I want to select which model runs in the subagent, because I know what subscriptions and limits I have.
- The future is a manager agent launching workers, but you still need to design that system: the rules, the permissions, the conditions when a subagent is launched. I don't want some guy at Anthropic or OpenAI deciding that for me.
ADRs.
- This is how you put decisions into a codebase. /docs/adr is one of the first folders I create in any project.
- Every core architectural decision gets a short file: what was decided, why, and what the state of the project was at the time. Some things can be read from code, but not everything.
- The stuff that can't should be documented, so future agents and humans instantly understand why it's built this way.
Tests.
- Current models BLOAT your repo with tests: unit tests, integration tests, database tests, even on the smallest repos where it makes no sense.
- If you tell the model to add tests, it adds an insane amount. If you tell it "don't add tests", it still adds some, and you land at the right amount.
Prod DB access.
- Any product with real usage needs to do this... create a read-only Postgres role and give your agents that.
- Don't give them write access. All it takes is one irreversible change and you're going to regret it.
- But no access is also a mistake. With read-only access, you can reality-check every feature. Does this even happen in production? Are people even using this? I wish I did this sooner.
Track your agentic productivity.
- We just released a new open-source repo under Vectal Labs called [agentic-productivity](https://github.com/vectal-labs/agentic-productivity). It tracks your commits, your agent sessions, and your user prompts.
- Each is a bad metric on its own, but combine the 3 and look at the long-term trend, and you can see whether you're actually becoming a better agentic engineer.
That's the setup as it currently stands. In a month, it's probably going to be different. This changes all the time.
by David Ondrej (spoken for YouTube, then re-written for article format)
## 相关链接
- [David Ondrej](https://x.com/DavidOndrej1)
- [@DavidOndrej1](https://x.com/DavidOndrej1)
- [130K](https://x.com/DavidOndrej1/status/2094424967345496191/analytics)
- [Link to my YouTube here.](https://www.youtube.com/@DavidOndrej)
- [pi.dev](https://pi.dev/)
- [github.com/davidondrej/skills](https://github.com/davidondrej/skills)
- [agentic-productivity](https://github.com/vectal-labs/agentic-productivity)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:59 PM · Aug 31, 2026](https://x.com/DavidOndrej1/status/2094424967345496191)
- [130.7K Views](https://x.com/DavidOndrej1/status/2094424967345496191/analytics)
- [View quotes](https://x.com/DavidOndrej1/status/2094424967345496191/quotes)
---
*导出时间: 2026/9/1 09:04:04*
---
## 中文翻译
# 智能体工程配置（2000+小时经验总结）
**作者**： David Ondrej
**日期**： 2026-08-31T13:59:59.000Z
**来源**： [https://x.com/DavidOndrej1/status/2094424967345496191](https://x.com/DavidOndrej1/status/2094424967345496191)
---
在过去 3 年里，我花在 AI 编程上的时间远超 2,000 小时，并且亲自采访了智能体工程领域最高效的一批人。
以下是我截至目前（2026 年第三季度）完整的智能体工程配置。
## 界面
这是指你与智能体交互的 UI / CLI。我的主要界面是 bb。
它是开源的、完全免费，让你能在同一个 GUI 中使用任何订阅、任何智能体、任何模型。Codex、Claude Code、Pi、Cursor CLI、OpenCode、Grok Build、Hermes，全部在同一个界面里。
Codex 或 Cursor 这类应用的问题在于，它们只允许使用自家的模型和订阅。而目标是花最少的钱获取最多的 token。
你在 Codex 或 Cursor 应用里喜欢的所有功能，bb 里都有，而且每周都在改进（而且它是完全开源、100% 免费的）。
我经常用的另一个工具是 cmux。
启动一个新的 cmux 工作区时，你可以像在 tmux 中那样分屏（这也是名字相似的原因），在每个窗格中打开不同的终端，还有内置浏览器。
cmux 的问题出现在当你有很多智能体和工作区时。左侧边栏真的不是合适的组织方式。同时处理几件事没问题，但对于大规模的严肃智能体工程工作，它不是最佳选择。
我用 Ghostty 作为终端，因为它非常快而且是原生应用。
在 Ghostty 里，你可以运行 Herdr，它基本上就是给智能体用的 tmux，是一个智能体后端运行时。非常精简、非常轻量、驻留在终端中，当某个智能体完成时，它的状态会显示在左侧：完成、空闲、阻塞、运行中。
追踪 AI 智能体的状态是至关重要的。我的预测是，在 3 到 6 个月内，这种“智能体状态追踪”会变得越来越重要，因为你不会再只跟一个智能体对话。你会跟一个管理大量工作者智能体的管理智能体对话。
最后一个必须提到的界面是 Corral，这是我自己开发的工具。
它不再是在智能体完成时随机切换（在 Herdr 中没有真正的顺序），而是让每个智能体都有一个优先级，就像任务有不同的优先级/重要性等级一样。当一个 P1 智能体完成时，它会置顶。当 P1 智能体完成运行时，你绝不应该去回复一个 P4 智能体。（是的……我确实需要把 corral 开源，还没来得及做。）
## 模型与订阅
你要用最少的钱获取最多的 token。这应该是每个智能体工程师的主要目标之一（在把活干完之后）。
目前有 4 个主要订阅，是的……两个月后这一切可能就完全不同了。
目前性价比最好的是 OpenCode Go。只要 $10，你就能用 Kimi K3、Grok 4.6、GLM 5.3、DeepSeek V4 Pro 等许多模型……但它没有最好的模型，比如 Fable 5 和 GPT-5.6 Sol（而且用量限制相当小）。
所以，如果你预算多一点，应该这样做：
- $30 —— OpenCode Go + ChatGPT Plus（$20）
- $50 —— 再加上 $20 的 Claude Code 订阅。
- $70 —— 再加 Cursor 每月 $20，你就集齐了所有订阅的最低档。
- $110 —— OpenCode + 一个大套餐。ChatGPT 的 $100 套餐比 Claude 划算。事实就是这样。OpenAI 算力更多，更愿意补贴。
- $210 —— 直接把两个 $100 套餐都买了。
- 如果你真的认真（像我一样），就把三个 $200 套餐全买了（Codex、Claude、Cursor），因为那些是 20 倍用量，这些套餐性价比最高。
顺便说……Cursor 套餐被严重低估了。Cursor，也就是 Grok，因为 SpaceX 收购将成为一个很棒的订阅。SpaceXAI 有海量算力，所以他们有资本玩补贴游戏。而且——我认为——Cursor/Grok 套餐给你 Cursor + Grok Bot 分开的用量额度，这太不可思议了。
Grok Bot 正迅速成为人们与智能体交互的新方式，所以拥有 Cursor 订阅从未如此重要（不是赞助哈哈，只是实话实说）。而且，新模型——Grok 4.7——也即将发布。
无论如何，千万不要付 API 定价的费用。那是市面上最亏的方案。直接买订阅。
## 云端智能体
很明显，云端智能体是未来。Cursor、Amp、Devin、Codex……这些公司都在云端智能体上押下重注。
证明云端智能体是未来的证据是下面这张图。
[image here]
这是 Cursor 内部来自云端智能体的已合并 PR 占比：今年年初约 10-15%，现在接近 60%。而且这是已合并的 PR，是真正被使用的部分。很快这会是 70%，然后 80%，然后 90%。
在本地机器上运行所有智能体的问题在于，它不可扩展。你不可能同时运行成百上千个智能体。只要几个智能体决定同时运行你的整个测试套件，你的电脑就会开始发出奇怪的声音（连我 $7,000 的 MacBook Pro 都吃不消）。
云端智能体给你隔离的环境、持久的会话、持续的互联网和电力供应。合上笔记本，你会丢失会话；断网几分钟，harness 就无法自我恢复。
现有云端智能体方案的问题在于极其严重的生态锁定。配置环境和所有密钥要花很多小时，然后你就被锁定了：会话在那里、你用他们的定价、你把所有数据都给他们。就算他们不用你的数据训练模型，也有很多其他方式使用你的数据。
解决方案是拥有自己的服务器。多亏了 AI，这只需要大约 10 分钟就能搞定（真的）。只需买一个 VPS，在上面运行 Herdr，然后 SSH 进去。
Herdr 给你持久的智能体会话，SSH 让你能从手机、笔记本等任何设备连接。你只需花几美元，就能实现云端智能体 80/20 的核心价值，而且没有锁定。
## 搭建你自己的云环境
我用 Hostinger 的 VPS，KVM2 套餐就足够了。我最想传达的是……你不需要是 VPS、DevOps、Linux 方面的专家，什么都不需要。
直接用大白话跟你的智能体对话就行了！！！
在即将发布的视频中，我现场搭建了整套环境。一个 cmux 工作区，左窗格是一个编程智能体（运行 Grok 4.6 的 Cursor CLI），右边是一个空终端窗格。
我的 cmux skill 让智能体找到另一个窗格并在其中执行命令。我自己先 SSH 进全新的 VPS，然后告诉智能体“了解这台服务器的一切，并搭好开发环境。Herdr、Node.js、Python 3、Git”。
它几秒钟内就分析了 VPS，安装了所有东西，启动了 Herdr，然后安装了 Pi Agent，在我的 MacBook 上找到一个 OpenRouter 密钥，自己完成了全部配置。短短几个提示词之后，我就让 Pi 运行 GPT-5.6 Sol，第二个会话运行 Fable，都在云端、在我自己的 VPS 上、拥有完整的 root 权限。
如果我的电脑或 Wi-Fi 出了问题……如果我的 MacBook 爆炸了，那些智能体仍会继续运行。完整教程在视频里。[我的 YouTube 链接在这里。](https://www.youtube.com/@DavidOndrej)
再提一个提速技巧……我用 SuperWhisper 来听写输入。正在读这篇文章的大多数人打字速度可能是每分钟 40 到 50 个词。那非常慢。
但是！你说话可以达到每分钟 250+ 词。语音 AI 工具（如 Superwhisper、Glaido、Whispr FLow）能立刻让你发提示词的速度提高 3-4 倍。用它。别犯傻。
## Harness
首先要提到的 harness 是 Pi Agent， Goat（史上最强）。
市面上最精简的 harness：只有 4 个工具，始终以 YOLO 模式运行，支持任何模型、任何提供商。非常优雅、超级可配置，这也是为什么这么多人在 Pi 之上构建。它是开源的、完全免费的，直接去 [pi.dev](//pi.dev) 获取。没得商量。它是我在 VPS 上安装的第一个 harness。
Cursor CLI。被严重低估，因为你可以使用所有模型：Grok、GPT 系列模型、Anthropic 模型、Kimi。可以打标签使用 skill，还可以预发送消息。总体上很棒的 harness。
下一类 harness 我称之为“自我改进”型 harness。
最流行的两个是 Hermes Agent 和 Prime Agent。这适合你不知道自己在做什么的时候。如果一个任务有很多不确定性、需要大量摸索，就使用自我改进型 harness，因为它会创建 skill 并随时间与你共同进步。
最后是经典款，Claude Code 和 Codex。我把它们设成了别名。很多人每天都敲 claude --dangerously-skip-permissions。极慢，极低效。我敲 cc 就能以跳过权限的方式启动 Claude Code；cx 以 YOLO 模式启动 Codex。
你必须为经常运行的长命令创建全局别名。这是智能体工程的法则之一：如何在同样的时间内完成更多的事？
## Skills（技能）
我的 skills 仓库上个月火了。（见 [github.com/davidondrej/skills](//github.com/davidondrej/skills)）它同样完全免费、开源，等等。
与智能体工程最相关的 SKILL 有：
*(1) /total-review
- 它会运行另外两个 skill，/gpt-review 和 /fable-review，分别用 GPT-5.6 Sol 和 Fable 5 审查你刚做的代码改动，然后把两个列表去重合并成一个只包含真正重要问题的列表。这就像请你最聪明的朋友们审查你的求职申请，他们只告诉你最大的问题。在中等到大型的改动上运行它，尤其是当由另一个模型完成时。如果工作是 Grok 4.6 做的，你会希望一个完全不同的模型来审查。
重要提示：任何你重复得足够多的操作都应该变成预设。
如果是单步操作，用文本替换。我把它们设成了 Raycast 代码片段。“用大白话简短回答。”“把你上一个回答改得更简单更短。”“暂存所有文件，写一条清晰的 commit，推送到 GitHub。”
如果是多步工作流，就把它变成一个 skill。
(2) /ask-then-build
- 我每天都在用这个 skill，在任何构建之前。与其说“让它兼容 Windows”然后让模型默默做出你以后可能后悔的重要架构选择，不如让它逐一带你过一遍主要决策，并提供选项。AI 模型擅长编码、擅长实现。它们没有品味，没有好的判断力。你作为人类，需要在这件事上保持主导。
(3) /deepapi
- 这个 skill 用于一切深度研究、一切爬取、一切网络相关的事。Codex 和 Claude Code 自带基础的网络搜索，但没有爬取、没有深度研究，而且它们很容易被屏蔽。“运行 8 次快速网络搜索，给我前 3 个选项”“爬取 Twitter”“爬取 GitHub”“找到 3 种联系某人的方式”。我团队里的每个人都在用它。必备。
(4) 护栏与推送锁
- 更无聊，但绝对必要，而且只需设置一次。全局智能体护栏是一个 pre-tool-call 钩子，确保你的智能体永远不会清空你的磁盘、永远不会覆盖 Git 历史、永远不会碰你的密码管理器。还有 push lock，用于当你并行运行 15 个以上智能体时：整艘船只有一个操作系统级的内核锁。合并、验证、推送、CI、部署、健康检查。
不要安装我所有的 skill。只拿你需要的。
## Worktree（工作树）
Worktree 基本上是把你的主检出复制到一个单独的文件夹，并在那里创建一个新的 Git 分支，这样智能体就能完全隔离地并行工作。
在小项目上，这完全是杀鸡用牛刀。留在单一分支上，工作得更快。
在需要时刻运行 20-30+ 个智能体的中大型项目上，这不可避免。没有 worktree，智能体会相互冲突、互相回滚对方的改动、互相打架。BB 的另一个优点：它内置了 worktree。它记得在我的大仓库上，我总是想要基于 origin/main 的新 worktree。
## 其他智能体工程技巧
知道什么时候该用哪个模型。
- 设计方案或启动新项目？用 Fable。它的灵光一现最多。修复一个深层、严重的 bug？用 GPT-5.6 Sol，最高推理强度。默认聊天？用高档位的 Grok 4.6。智力几乎一样，但便宜 2 倍、快 2 倍。前端？用 Kimi K3。
- 当一个重要的新模型发布时，留出一天只用那个模型。别听 Twitter 上的。自己试试。
知道什么时候该审查。
- 我不会对每个改动都跑 total review。一个小前端微调直接上生产。
- 绝对不要递归审查。如果你告诉一个模型“找出 5 个最大的问题”，即使代码库完全没问题，它也会找出 5 个问题。这些模型会编造想象中的 bug。
预发送。
- 通常我知道智能体接下来会做什么，所以我会提前排队消息：“实现这个计划”“对这个跑 Fable 审查”“现在修复那些问题”。
- 有时是 2 条消息，有时是 6 条。
- 永远不要用不允许预发送的 harness。
子智能体被过度使用了。
- 很多人只是在它们身上烧光自己的额度。我在可控的时候才用它们。我想选择哪个模型在子智能体中运行，因为我了解自己的订阅和额度情况。
- 未来是管理智能体启动工作者智能体，但你仍然需要设计那个系统：规则、权限、什么条件下启动子智能体。我不想让 Anthropic 或 OpenAI 的某个人替我做这个决定。
ADR（架构决策记录）。
- 这是你把决策沉淀进代码库的方式。/docs/adr 是我在任何项目中最早创建的文件夹之一。
- 每个核心架构决策都有一个简短的文件：决定了什么、为什么、当时项目的状态是什么。有些东西可以从代码中读出来，但不是所有。
- 读不出来的就应该记录下来，这样未来的智能体和人类就能立刻理解为什么这样构建。
测试。
- 现在的模型会用测试让你的仓库臃肿不堪：单元测试、集成测试、数据库测试，甚至在最不需要测试的小仓库上也是如此。
- 如果你让模型添加测试，它会加多得离谱的数量。如果你告诉它“不要加测试”，它还是会加一些，最后你反而会落到刚刚好的数量。
生产数据库访问。
- 任何有真实使用量的产品都需要这样做……创建一个只读的 Postgres 角色并把它给你的智能体。
- 不要给它们写权限。只需要一次不可逆的改动，你就会后悔。
- 但完全没有访问权限也是错的。有了只读权限，你可以对每个功能进行现实检验。这在生产环境中真的发生了吗？真的有人在用吗？我真希望自己早点这么做。
追踪你的智能体生产力。
- 我们刚在 Vectal Labs 下发布了一个新的开源仓库，叫 [agentic-productivity](https://github.com/vectal-labs/agentic-productivity)。它追踪你的提交、你的智能体会话、以及你的用户提示词。
- 每一项单独看都是糟糕的指标，但把这三者结合起来看长期趋势，你就能看出自己是否真的在成为更好的智能体工程师。
这就是我目前的配置。一个月后，它可能就不一样了。这东西一直在变。
作者：David Ondrej（先为 YouTube 口述，后改写为文章格式）
## 相关链接
- [David Ondrej](https://x.com/DavidOndrej1)
- [@DavidOndrej1](https://x.com/DavidOndrej1)
- [130K](https://x.com/DavidOndrej1/status/2094424967345496191/analytics)
- [我的 YouTube 链接在这里。](https://www.youtube.com/@DavidOndrej)
- [pi.dev](https://pi.dev/)
- [github.com/davidondrej/skills](https://github.com/davidondrej/skills)
- [agentic-productivity](https://github.com/vectal-labs/agentic-productivity)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [2026年8月31日 晚上9:59](https://x.com/DavidOndrej1/status/2094424967345496191)
- [13.07万 次浏览](https://x.com/DavidOndrej1/status/2094424967345496191/analytics)
- [查看引用](https://x.com/DavidOndrej1/status/2094424967345496191/quotes)
---
*导出时间: 2026/9/1 09:04:04* 
