↑ 章节地图:课程索引:01-介绍 · v0 · v1 · v2 · v3 · v4 · v5 · v6 · v7 · v8 · v9 · v10 · v11 · v12 · v13

IndyDevDan · 51:36 · 发布 2026-02-23 · 22.8万次观看(截至抓取) · 观看原视频 🔥 观众最高回放 29:56 — 本片总论点(why Pi 反主流)
˛≈≈
⚠️ 本片作者章节标题与实际口播在某些版本编号处滞后(主会话记录到差异),各章笔记按实际口播内容整理,章节标题保留作者原标。
主旨
一句话:Pi(Mario Zechner 出)是开源 AI 编码 agent,定位是 Claude Code 的开源 + 完全可定制 + 无安全 theatre + 模型不绑定反主流选项;视频演示 14 种实战定制版,从心流 minimal 到多 agent 编排再到 meta-agent,借 80/20 策略(80% Claude Code + 20% Pi)做对冲。
核心论点
- Claude Code 出现 cancer——增长+利润驱动使其从为工程师定制偏离为服务大众,你能动手的部分被反复抽象,只剩 harness 这一层你还能控制。(→ 详解见 01 章哲学部分)
- Pi 的两个一贯主张:开源 + 全可定制;与 Claude Code 的好用但封闭形成产品哲学对位。
- Pi 工程哲学四对照:200 token 系统提示 vs Claude Code 10000 token、YOLO mode vs 五种审批、底层完全观察 vs 抽象掉、任意模型 vs 优先自家。
- 三 tier 实操路线:Tier-1 基础 harness(v0-v6) / Tier-2 多 agent 编排(v9/v12) / Tier-3 meta-agent(v13);每章演示一种或多种 PI 变体,逐步把 agent 的可定制面撑开。
- 80/20 策略:80% 任务仍用 Claude Code 取开箱体验,20% 用 Pi 取 harness 全控制+多 agent + 实验场,思路从 pick one 换成 think in ands。
- 限定 Pi 适用于知道自己在做什么的进阶工程师(中高级+),YOLO 默认与全定制面组合下不推荐 vibe coder 直上。
章节地图
- 第 1 章 · 介绍:PI Agent vs Claude Code — 01-介绍.md:Pi 的产品哲学与设计理念,把开源 + 全可定制 + 无安全 theatre + 多模型立成对 Claude Code 的反主流选项,引出三 tier 路线。
- 第 2 章 · PI v0 Default PI — 02-v0-Default-PI.md:原样启动 Pi:200 token 系统提示、四默认工具、YOLO 模式、thinking 暴露的标准开箱体验。
- 第 3 章 · PI v1 Pure Focus PI — 03-v1-Pure-Focus.md:把 footer 全去掉、只剩一栏纯输入,牺牲可见性换心流。演示 Pi 顶部 UI 可被整层剥离。
- 第 4 章 · PI v2 Minimal PI — 04-v2-Minimal.md:裁到 bare minimum:read/write/edit/bash + 自建 task 工具,会话级 tool 仍可被覆盖,Claude Code 默认工具数一半。
- 第 5 章 · PI v3 Cross Agent PI — 05-v3-Cross-Agent.md:多个 location 全局/项目级 skills/commands/agents 一次性全部 load,Pi 让你声明工具从哪里读。
- 第 6 章 · PI v4 Purpose Gate PI — 06-v4-Purpose-Gate.md:开 agent 时先弹问这 agent 的目的是什么,把回答追加到默认 200 token 系统提示,自建 widget 持续在屏。
- 第 7 章 · PI v5 Tool Counter PI — 07-v5-Tool-Counter.md:在 footer 计 context + model + 项目分支之外,加一个等待工具的状态行。扩展 footer 给自定义信号留位。
- 第 8 章 · PI v6 Tool Counter Widget PI — 08-v6-Tool-Counter-Widget.md:不只 footer,起一个独立 widget 在底部累计工具调用次数 + 名称 + 主题循环。模板化扩展即组件。
- 第 9 章 · PI v7 Subagent Widget PI — 09-v7-Subagent-Widget.md:Pi 没有原生 subagent,用生命周期钩子自造一个 widget 子代理面板,/sub + 自然语言即起例,可并行与回收。
- 第 10 章 · PI v8 Tilldone PI — 10-v8-Tilldone.md:未完不许走:钩子拦 ls 必须先建任务清单,Haiku 都能跑完。把任务即会话做进 harness 的确定性。
- 第 11 章 · PI v9 Agent Team PI — 11-v9-Agent-Team.md:scout/plan/build/review 的 agent team YAML 配置项 + /agent-team 切换,primary agent 改当 dispatcher 不动手。
- 第 12 章 · PI v10 System Select PI — 12-v10-System-Select.md:/system 改写当前会话的 system prompt,随时把 Pi 换成 browser agent / planner 等,会话级 persona 切换。
- 第 13 章 · PI v11 Damage Control PI — 13-v11-Damage-Control.md:明确在 hooks 层禁掉 rm -rf 之类高危命令。把所有安全 theatre 做在一个 extension 里。
- 第 14 章 · PI v12 Agent Chain PI — 14-v12-Agent-Chain.md:3 个 scouter 串成 agent chain / pipeline,前一 agent 的输出直接喂下一个,plan→build→review 是其中一种。
- 第 15 章 · PI v13 Meta PI — 15-v13-Meta-PI.md:8 个领域专家 agent 各自专精 Pi 的某一面,meta 主 agent 并行调并拿总论。meta 是造 agent 的 agent。
价值定位
对该视频目标学习者的价值:
- 适合谁:已在用 Claude Code 想要在 harness 层拿到完全控制权、或对多 agent 编排(plan→build→review)有强烈定制需求的工程师
- 解决什么:14 种实战 PI 变体覆盖了心流 minimal → 状态行/widget → subagent + tilldone loop → agent team/chain → meta-agent 全光谱定制操作
- 认知 vs 实操:介于二者之间,以工程演示带认知(为什么 harness 这一层值得投入)+ 大量可照搬的 TypeScript extension snippet
- 与 Harness、Harness 大师课、Omnigent 重叠:这些笔记已定调 harness 占 90%,本片是具体到 Pi 这一开源实现的 14 种可执行样板
立场与利益(整体分档)
- 与利益同向(待印证):harness > 模型 — 与 cole-medin / chase-ai 多源同向,有商业卖课嫌疑,但主张可独立印证
- 利益中性:14 种 extension 演示、Pi/Claude Code 字段对照表 — 通用工程演示,无直接挂商品
- 与利益反向(单独标):作者主动说出 Pi 不支持 subagent / MCP / 多 agent 原生功能,所以你必须自己造,反向印证工具未到位以换取”可造”卖点,损商业形象
视频 description 含导流:https://agenticengineer.com/tactical-agentic-coding?y=f8cfH5XX-XU(带 ref 参数的付费课链接),与”harness 那 90%“主张同向,采信前需对照独立信源(Harness.md 三源已印证)。
关联
- 印证(独立来源共识 harness 是 90%):Harness、Harness 大师课 — 三信源共识 harness 是能动工程可优化层
- 互补(具体到 Pi 这一实现):Omnigent — 那是多 agent 编排元层实现;本片 tier-2 是 Pi 自己接 TypeScript extension 实现,粒度更细
- 进阶(同主题深一层):CMUX 实测 — CMUX 解决”看不见起不动的 agent”的可视化批管理;本片 v9 是 YAML agent team 自定义
- 互补(对照 Claude Code 设计哲学):CC 高阶课 · 并行化基础 — Pi 与 Claude Code 在 subagent 哲学上对照(原生 vs 自建)
自检问题(看完本课程)
- Pi 的产品哲学哪三个反主流立场?请点名与 Claude Code 的对照。 答案:立场三件套——200 token 系统提示、YOLO mode、底层全观察;对照见 01 章的哲学四对照。回跳 00:00
- 三 tier(基础 harness / 多 agent 编排 / meta-agent)如何映射到具体 PI 版本号? 答案:Tier-1 v0-v6 + v10(基础 harness / 状态栏 / widget);Tier-2 v9 + v12(Agent Team / Chain);Tier-3 v13(Meta-PI)。回跳 02:00
- 作者建议的 80/20 策略具体怎么用? 答案:80% Claude Code 取开箱体验,20% Pi 取 harness 全控 + 多 agent + 实验场,思路换成 think in ands。回跳 01:51
💬 热门评论 top-20 主 + 16 回(抓取 2026-07-07)
[2] @brucemvanhorn2:他们借鉴了美国海军陆战队的《步枪手信条》:这把步枪是我的。像这样的枪有很多,但这一把是我的。我的步枪是我最好的朋友。它是我的生命。我必须掌握它,就像我必须掌握我的生命一样…… 👍 87 ↪ @christopher-s4d7d:因经典电影《全金属外壳》而流行起来。 👍 6 [3] @stevenpeterson1031:Pi 真是一股清流,很高兴它被提到了。我用更少的 token 完成了更多的工作。现在的框架太臃肿了,模型已经足够聪明,给一些基本工具就能搞定剩下的,不需要塞一堆适得其反的上下文。 👍 16 [5] @ed-fruty:下一期视频:Pi 对比 OpenCode。 👍 266 ↪ @yashnu:OC 太棒了。 👍 22 ↪ @kyconnol:nvim - Pi LazyVim - OpenCode 两者都很棒! 👍 8 ↪ @danielmoll:我两个都留着。我不是 swarm agents 的粉丝——opencode 里的规划功能很方便,在同一个文件夹里切换两个工具也挺好。 👍 7 ↪ @ugoenyioha3201:我在看这个视频,没看出和 opencode 有什么本质区别……可能是因为我给 opencode 写了个 agentic teams 移植版,这些功能我早就有了。很多功能用 oc 插件或技能也能实现。不过我喜欢它的模块化。是用 Rust 还是 Go 写的?我会去看看。我喜欢它的模块化,会去试试。 👍 5 ↪ @bzbetty1:喜欢 opencode,但除非他们加入 hooks(他们有插件,但配置起来没那么方便),不然我觉得 Dan 不会用。 👍 2 [6] @FiatSceptic:一开始我很喜欢这个想法,但加完所有我需要的东西后,我发现我基本上重建了 opencode 的功能,只是 UI 略有不同。 👍 52 ↪ @abj136:我有点没听懂你的意思。你是说 Pi 不完整,当你补上缺失的部分后,基本上就是 OpenCode 了。 👍 4 ↪ @HanzHermannHoppe:那你不如直接用 opencode,但每个人的用例不一样。这就是 Pi 的好处,你可以把它做成你需要的,不多不少。比如我,我需要通过语音命令远程控制会话,Pi 的 RPC 接口让我很容易就写了一个最小功能的 app 来满足这个需求。 👍 1 ↪ @jkf16m96:我理解这个观点,但同时你也可以往回退一点,变得更简单。比如,我一直在打磨小语言模型的工作流。我尝试使用 Pi 子代理的扩展。另外,技能也很不错。Pi 的自我意识和可扩展性让社区很容易做出扩展,这些扩展能感知模型上下文窗口的限制。这非常有价值,很多人已经用更复杂的模型做到了同样的事情。 ↪ @pasemes:这基本上就是我在人们用了 IntelliJ IDEA 或 Visual Studio 之后再用 VS Code 时的感受。 ↪ @oceanheartai:一段有教育意义但最终可能令人失望的旅程…… 👍 5 [7] @CodveTools:对 Pi 的剖析很棒!token 效率的说法很有说服力——当模型足够聪明,能用最少的上下文解决问题时,那些臃肿的框架就显得多余了。希望能看到后续视频,对比 Pi 和 Claude Code 在处理多文件重构任务时的表现。🔥 👍 5 [10] @superresistant0:哈哈,你终于用 Pi 了!!!我早就说过 Claude Code 太臃肿了,说个 hello 就要 20K 上下文。 👍 74 ↪ @Chensabuaka:确认了,测试用了 18k。 👍 5 ↪ @SolusArcticWolf:不过你还是需要单独的订阅。你可以调整 token 数。 👍 1 ↪ @balajiramasamy3863:能分享一下你的 Pi 编码代理设置吗?我试过,太简陋了,连权限检查都没搞通。 [14] @TaylorsTechLog:请多做关于 Pi 的视频。特别是你怎么在系统层面设置它,然后为单个项目配置。好视频。 👍 1 [16] @cgintrallying:pi-coding-agent 是你需要的稳定伙伴。感谢精彩的讲解——我用了很多 Pi,但还没碰过扩展。 👍 4 [17] @mapledev9335:太棒了 Dan。这个视频里干货真多。我在我的 OpenClaw 机器人上安装了 Pi,然后修改了 GSD 仓库(Get-Sh!t-Done),让它作为 Pi 的扩展而不是 Claude Code 的扩展运行。我的 OpenClaw 机器人的工作质量提升了 10 倍,而且我只跑了 qwen3.5:397b。我的 pi-gsd 扩展还在开发中,但到目前为止结果让我印象深刻。 ——其他 11 条:感谢/夸赞([1]@aaaa-f8g6s、[4]@basically335、[8]@uhtexercises、[9]@robertpastorella、[11]@sorenfrederiksen155、[12]@LinardsBerzins、[13]@unsupervised-learning、[15]@inbar86、[18]@wires__、[19]@Spryax、[20]@AMR2442)
英文原文
[1] @aaaa-f8g6s:youtube algorithm has called me and here i am 👍 18
[2] @brucemvanhorn2:They're borrowing from The Rifleman's Creed from the US Marine Corps: This rifle is mine. There are many like it, but this one is mine. My rifle is my best friend. It is my life. I must master it as I must master my life... 👍 87
↪ @christopher-s4d7d:Popularized by the classic Full Metal Jacket 👍 6
[3] @stevenpeterson1031:Pi is a breath of fresh air, i'm glad it's getting some mention. I consistently get more done with fewer tokens. Modern harnesses are becoming overloaded when the models are more than smart enough now to be handed some basic tools and figure out the rest without a bunch of counterproductive context stuffing. 👍 16
[4] @basically335:Legend, I was just integrating oh-my-opencode into my AOC IDE but this is exactly what AOC needed 👍 4
[5] @ed-fruty:Next video PI vs OpenCode 👍 266
↪ @yashnu:OC Rocks. 👍 22
↪ @kyconnol:nvim - Pi LazyVim - OpenCode Both are great! 👍 8
↪ @danielmoll:i keep them both. i am not a swarm agents guy - planing in opencode is so handy and switching in the same folder with 2 tools is nice :) 👍 7
↪ @ugoenyioha3201:I’m watching this. And I’m not really seeing any unique differentiation from open code…..perhaps because I wrote an agentic teams port for opencode and have this capability at my fingertips. You can do a lot of this with oc plugins or skills. I like the modularity though. Is it written with rust or go? Will check it out. I like the modularity though. Will check it out. 👍 5
↪ @bzbetty1:love opencode but unless they add in hooks (they have plugins but not as easily configured) I can't see Dan going for it. 👍 2
[6] @FiatSceptic:I loved the idea at first but after I've added everything I needed I came to the conclusion that I've rebuild the same that opencode offers with slight UI difference 👍 52
↪ @abj136:I struggled to understand you. You’re saying Pi is incomplete, and when you fill the missing bits you basically have OpenCode. 👍 4
↪ @HanzHermannHoppe:then you're better off using opencode, but not everybody's use-case is the same. That's the nice thing about pi, you make it into what you need and nothing more. Me for example, I needed a way to remote control sessions via voice commands and pi's RPC interface made it easy to code a minimally functional app to meet this need. 👍 1
↪ @jkf16m96:I understand this view but at the same time. You can try to backtrack a little, to go simpler. For example, I've been polishing a workflow for small language models. I've been trying to use the extension of pi subagents. Also, skills are very nice to have. Having pi being as self aware and extensible, makes it easy for the community to make extensions that are aware of the limits of the context window of models. This is extremely valuable, a lot of people have been able to build stuff that more complex models are able to do too.
↪ @pasemes:That is essentially what I felt when people were using VS Code after IntelliJ IDEA or Visual Studio
↪ @oceanheartai:An educating but perhaps ultimately disappointing journey... 👍 5
[7] @CodveTools:Great breakdown on Pi! The token efficiency argument is compelling - when models are smart enough to figure things out with minimal context, the heavy harness frameworks start to feel like overkill. Would love to see a follow-up on how Pi handles multi-file refactoring tasks compared to Claude Code. 🔥 👍 5
[8] @uhtexercises:Knew that would come. Was waiting for it. Natural progression.Thanks. 👍 4
[9] @robertpastorella:Love the format of these videos and the content as well! Good on ya!
[10] @superresistant0:LOL finally you're using Pi !!! I told you claude code is f'ing bloated 20K context to say hello. 👍 74
↪ @Chensabuaka:confirmed. 18k for test 👍 5
↪ @SolusArcticWolf:Still you need separate subscription. You can tune tokens. 👍 1
↪ @balajiramasamy3863:Can you share your Pi coding agent setup? I tried using it. It's very bare bone and I couldn't even get their permission check working.
[11] @sorenfrederiksen155:Thanks for posting this. I haven't tried this yet, but will clone as soon as I get back to the computer. 👍 1
[12] @LinardsBerzins:Hope this will be added eventually to your course. Im building basic skills first to take your course eventually. Amazing channel and course 👍 2
[13] @unsupervised-learning:Oh man, got so excited every time he said "Pie". Thought he was calling us out! : ) But ours is a different PAI. lol Great video as always. And maybe Pi will be the base underneath PAI sometime in the future! 👍 30
↪ @sblacksg:We need PAI free from Claude Code, PAI with PI! 👍 4
↪ @nikolagavric6812:My thought when i started watching this video :) 👍 1
[14] @TaylorsTechLog:Please do more on Pi. Specifically how you set it up system wide and then configure for individual projects. Great video. 👍 1
[15] @inbar86:This is super cool! I'm going to start using it from now on. Also very cool video - what did u use to create it? 👍 2
[16] @cgintrallying:pi-coding-agent is the stable partner you need. Thanks for the awesome walkthrough - been working with pi a lot but not touching extensions yet . 👍 4
[17] @mapledev9335:Awesome Dan. There's so many goodies in this video. I installed pi on my OpenClaw bot and then modified the GSD repo (Get-Sh!t-Done) to work as an extension in Pi instead of Claude Code. The quality of work I'm getting out my OpenClaw bot has gone up 10x, and I'm only running qwen3.5:397b. My pi-gsd extension is a work-in-progress but so far I'm impressed with the results.
[18] @wires__:Your videos are so insightful. I learn so much that I really never would have otherwise. Thank you for making them! 👍 1
[19] @Spryax:Just hyped your channel thank you for the awesome info!
[20] @AMR2442:thank you for sharing this knowledge. I’m not even half way through the video 👍 3