开发者周报 Builder Weekly

OpenWorker AI that gets your everyday tasks done. OpenWorker is an open-source AI coworker that lives on your desktop and delivers finished work, not just chat: a polished document, a Slack reply with the numbers, an updated calendar, a triaged inbox.

How I Cut an AI Agent’s Token Use by 94% 我将该技能“编译”成了一个专用工具。这个新技能几乎算不上是一个技能了——它变成了一个调用 Python 程序的精简引导程序。该程序负责获取已知来源、构建清单、检查近期帖子、应用过滤器并管理工作流。它仅在选择和生成内容时才调用大语言模型。 https://tokenshrinker.com/

Programming

把“今日头条”式排序能力，放回你自己的 RSS 信源里 https://github.com/Pls-1q43/Dibao

省 Token 技巧：很多时候你根本不需要 LLM

Scikit-Learn Meets Large Language Models Scikit-LLM

从一条运行轨迹开始，一步步实现 Pi。

What I learned building an opinionated and minimal coding agent PI { Mario Zechner }

Agent swarms and the new model economics - cursor

Replit 如何規模化評測和持續改進 Vibe coding

ResearchStudio - microsoft is a collection of skills covering the entire research lifecycle — from an under-specified research direction to a published paper. The skills run on Claude Code and Codex.

微软开源 AI 科研神器：ResearchStudio-Idea 筛选可靠创新点

7千字看本体自动生成与执行实现：原型开源项目OntologyAutoGen_HW细节 https://github.com/whyseu/OntologyAutoGen_HW

5 Trends That Defined AI Engineering at World’s Fair 2026

12-Factor Agents - Principles for building reliable LLM applications

Loop Engineering with Adaptive PDF Parsing: Start Cheap, Pay for a Heavier Parser Only When the Page Needs It

Compare Tavily, Perplexity API, Google Search Grounding, Exa with LLM-as-Judge in LangSmith

Automated Data Analyst (ADA) — Open-source AI for CSV & Excel

I Built an AI Data analyst that tells you when it hallucinates

Best AI Skills for Research: 6 Skills to Run the Full Academic Workflow (2026)

Kill AI Slop Vibe-coded products are ugly in the same specific way: indigo gradients, glowing cards, emoji everywhere, a mascot in every corner, ALL-CAPS stat-cards. It’s what a machine reaches for when it has no taste but wants to look impressive, and it’s so common you’ve stopped seeing it.

kumo - Cloudflare’s component library for building modern web applications. Kumo provides accessible, design-system-compliant UI components built on Base UI.

Nimbus - Cloudflare Docs sites you own, where humans and agents are both first-class.

Others

体感 - 槽边往事

Garmin 对抗 AI 教练风潮？收购背后的商业逻辑

ASD-STE100 - Simplified Technical English Standard for Technical Documentation European Union Trade Mark No. 017966390. ASD-STE100 Simplified Technical English (STE for short) is a controlled natural language and an international standard to write technical documentation. It is fully owned by ASD, Aerospace, Security and Defence Industries Association of Europe, Brussels, Belgium.

市场是检验AI应用的唯一标准之一 “他们的应用之所以那么糟糕，一个原因就是他们没有场景，没有knowhow，也没有一些基本的商业意识。”

Hallmark A design skill for Claude Code, Cursor, and Codex that refuses to look AI-generated.

oh-story-claudecode 网文写作 skill 包，覆盖长篇与短篇网络小说的扫榜、拆文、写作、去AI味、封面图全流程。

向幸福投降

GC Minimal Zine Poster Codex skill for turning a theme, sentence, object, mood, article idea, photo, or content brief into a quiet minimal zine-style editorial poster prompt and a generated raster image.

The Naturalist’s Library A complete digital edition of William Jardine’s 40-volume series on natural history published 1833–1843, including more than 1,300 restored hand-drawn illustrations of birds, mammals, insects, and fish.

Dev weekly 2026-Jul-20

Sun, 19 Jul 2026 02:02:03 GMT

AI

NousResearch Accelerated Business Hackathon

Andrew Ng just dropped a 3-hour course on how to become an AI Engineer in 2026

Sutskever 30 implementations Comprehensive toy implementations of the 30 foundational papers recommended by Ilya Sutskever

《深入理解 AI Agent：设计原理与工程实践》李博杰

The map of AI.

中文机器学习研究线上课

用 LLM + 语义聚类，把海量用户评论提炼成四级 VOC 标签体系

DSLs Enable Reliable Use of LLMs

Programming

Telegram Serverless

GPT 5.6 Has 72 Possible Configurations. What’s A Good Default? GPT 5.6 中由不同模型规模与推理期算力组合而成的 72 种配置，并分析了如何基于训练期与推理期的缩放法则（Scaling Laws）在其中寻找最佳的性能与成本权衡

What building Shippy taught us about building agents

让 Agent拥有「跨终端长期记忆」——基于 Amazon Bedrock AgentCore Memory 的实践

Reverse Engineering ChatGPT Web: How OpenAI Built for a Billion Users

OpenWiki 0.2 is adopting the OKF support

weekly AI news digest with GitHub Agentic Workflows

ReactBench v1 an evaluation for coding agents on realistic React work

Agentic Testing: Where Agents Fit in the E2E Testing Stack 传统端到端测试强制执行固定的 UI 路径，而基于 AI 代理（Agentic）的测试则通过动态探索来验证目标。尽管代理测试在执行成本、性能和复杂场景下的稳定性方面存在权衡（Trade-offs），但它们作为更高层级的探索与调试工具，能够有效补充而非替代现有的确定性 CI 测试流水线。

The New Software Lifecycle

Memex 本地优先、具备类型系统，服务于人与 Agent 的知识工作空间

sail - Drop-in Apache Spark replacement written in Rust

拨开迷雾：炒得火热的FDE究竟是什么？九问九答剖析其是否自带银弹？

Turn your idea into an app

Open-source Reflect rewrite A local-first AI agent-friendly Markdown note-taking app

Five studies changing how I think about AI in software engineering

Others

Yamanote line sound

Agentic orchestration: Enterprise AI organizations have a deployment problem, not a platform problem — and most are calling chatbots agents

Enterprise AI is entering an evaluation gap: Agents are gaining autonomy faster than companies can verify them

Dev weekly 2026-Jul-13

Sun, 12 Jul 2026 02:02:03 GMT

AI

Harness Engineering for Self-Improvement 利用工程化手段实现自我提升

Better Models: Worse Tools 较新的 Claude 模型有时会在嵌套的 edits[] 数组中调用 Pi 的编辑工具时，加入额外且凭空捏造的字段。

文译专注于将多语言 EPUB、FB2 或 TXT 小说翻译成中文，并尽量保留 EPUB 原排版、图片、目录和跳转。

Getting started with loops

How Schneider Electric Built Their LLMOps Foundations At Enterprise Scale With LangSmith langsmith architecture

正式开源！美团 LongCat-2.0 同步开放国产卡推理代码 https://github.com/meituan-longcat/LongCat-2.0

LiteRT.js, Google’s high performance Web AI Inference

Tau A small, readable terminal coding agent — and a working example of how coding agents are built, from huggingface.

Improving Agents is a Data Mining Problem langchain

Gemma 4 Technical Report

Fine-Tuning Financial LLMs for Customer Intent and Experience 中文演说。Jui Sheng Chang and Walter Hsieh from Cathay Financial Holdings and Cathay United Bank will share how financial LLMs can be built and fine-tuned for customer intent classification and improved customer experience.

Machine Learning Systems 人工智能系统工程的原理与实践

Architecture 2.0

Recreate Our 1-Minute Seedance 4K Film - The Full Seedance 4K Breakdown With All Prompts 重现我们那部 1 分钟的 Seedance 4K 短片——Seedance 4K 全流程解析及所有提示词

The State of the AI Economy

Radar Trends to Watch: July 2026 值得关注的雷达趋势：2026 年 7 月

欢迎来到 Learn Harness Engineering

CS294/194-280 Advanced Large Language Model Agents

Agentic AI Architecture infoq电子书

Achieving operational excellence with AI

Fable is not a useful model

Pi: The Coding Agent That Hides Nothing

LLM-as-a-Verifier A General-Purpose Verification Framework

The new GPT-5.6 family: Luna, Terra, Sol

Programming

Announcing TypeScript 7.0

Rewriting Bun in Rust

Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase 模型所依托的工具框架对成本和质量有着巨大影响。在许多情况下，像 Pi 这样简单的框架在我们的工作负载中表现最好。

How Prompt Tuning Improved GPT-5.5 in VS Code

Teaching the Agent Our Craft: Structured Agentic Development on a Real Codebase

astryx - design system

SentencePiece 文本分词工具。快，轻，不用人工标注。给神经网络生成系统用

AgentHarness Evaluation harness for Apodex-1.0 on public deep-research benchmarks.

MarketFish - Don’t guess. Simulate 多智能体市场模拟引擎。不是问一个大模型”这个产品能行吗”，而是建造一个由 128+ 个 AI 消费者组成的数字市场——每个人都有独立的身份、预算、情绪和偏见——让他们在 30 轮中自由买卖。他们的购买决策、流失模式、社交影响，揭示真实用户会怎么做。

Maka 桌面 AI 工作台基于pi

Learning to Replicate Expert Judgment in Financial Tasks 总而言之，LLM 在资金调配中表现出高信念、叙事驱动的自由交易风格。结果更多由集中度、叙事执念和不对称下行风险决定，而非交易层面的优化。

Palantir Ontology的Neo4j+业务规则引擎平替实现及其如何”驱动”Agent？

Learning to Replicate Expert Judgment in Financial Tasks 学习在金融任务中复刻专家判断我们测试过的前沿模型在处理相对简单的金融任务时表现吃力，且模型本身的进步并未带来多少性能提升。相比之下，我们已经证明，使用由专家投资者标注的高质量专属数据集进行微调，可以产生在我们的任务上超越前沿性能的定制模型。我们发现，这一结果不仅适用于我们在本文中讨论的六项任务，在更广泛的领域也同样成立。

如何萃取老師傅的知識做成 Agent Skill? 六條路線的方法論框架

你的 Prompt 通得過 The Mom Test 嗎? 如何避免 LLM 的迎合問題

What is the future of software engineering with Adam Bender, Principal Software Engineer at Google

Six Things Enterprise Architects Get Wrong About the Operational–Analytical Chasm — and How to Close the Agentic Loop Snowflake

UML — See Your Codebase Point it at a folder. Get a diagram back.

Ten Years Away: Designing an Interactive Comic for Studio375’s Tenth Anniversary The story behind “Ten Years Away”, a scroll-driven interactive comic created to celebrate Studio375’s tenth anniversary, from its printed origins to its WebGL implementation.

Your Kafka Infrastructure Just Became Optional Datastream：兼容 Kafka 的原生端点

JavaScript in-page GUI agent. Control web interfaces with natural language.

Anomaly detection using dynamic thresholds and two-year-long alerts in Cloud Monitoring

CS25: Transformers United V6

Stanford CS25: Transformers United V6 I From Language Models to Native Multimodal Intelligence

Evaluating ChatGPT as a Portfolio Decision-Maker in Micro-Cap Equities

Heads of Enterprise Architecture Must Rewire EA for AI

Flint: A Visualization Language for the AI Era A Microsoft Research project

Elastic 开源了基于认知科学的 Atlas Agent Memory https://www.elastic.co/search-labs/blog/agent-memory-elasticsearch

Survey Says: Agentic AI Penetrates the Enterprise, but Some ROI Questions Remain 《Battery Ventures 企业技术支出状况调查》。该调查涵盖了 100 位资深技术领导者的反馈，他们所代表的年度技术支出总额超过 660 亿美元

Your LLM Has a Spending Problem - How Prompts, Context Windows, and Agent Loops Quietly Torch Your Budget 你的大模型（LLM）有“乱花钱”的问题——提示词、上下文窗口和智能体循环是如何悄悄掏空你的预算的 AIFinops

Others

AI 季报 26Q2：从 coding 到 RSI，强者愈强的未来？

播客｜从「西印度毁灭述略」到冷战核危机：漫谈古巴与加勒比海五百年

《晋书》专题｜林昌丈：魏晋史上的南与北

Product Design Psychology by Wouter de Bres Understand the minds you’re designing for and the mind you’re designing with.

iPhone Photography Awards - The 2026 Winners

mobile01 旅遊美食

浓雾与犯罪华生的回忆录里不止有风和雨（福华的秋窗风雨夕》，还有伦敦那尽人皆知的雾。

How to ask for help from people who don’t know you

宏观经济学 7 天速成课教学实验复盘报告

播客｜与柏琳漫谈巴尔干千年史：从古南斯拉夫诸国到近代科索沃危机

朱雀仿宋/朱雀宋朝「朱雀仿宋」是璇玑造字的开源仿宋字体计划

ML Job Interviews: The Ultimate Guide

RED FLAG - LOCOMOTIVE ACT 1865 1865年的英国，针对道路上日益增多的自行驱动车辆，立法做出的回应是《机车法案》（有时被称为《红旗法案》）。该法案包含多项条款，其中规定自行驱动车辆前方必须有一名步行者，手持红旗，以提醒其他道路使用者车辆即将到来。既得利益者的目的是阻碍进步，以保护他们在马匹、马车和火车上的投资。

Dev weekly 2026-Jul-06

Sun, 05 Jul 2026 02:02:03 GMT

AI

Build a Reasoning Model (From Scratch)

fast.ai—Making neural nets uncool again

Scaling Laws, Carefully lilianweng

LongCat-2.0 正式发布

The Hitchhiker’s Guide to Agentic AI: From Foundations to Systems

Same prompt, different models: Cowork model outcome and cost comparison

Guidelines for Respectful Use of AI

11-768: AI Agents Fall 2026 @ Carnegie Mellon University

Getting started with loops claude

MinerU 团队新产品 | Sciverse 科学数据基座，把可信的科学知识接入你的 AI 工作流

OCR’ing 100,000 pages with open-source VLMs on Modal redspring

Agent Evaluation Readiness Checklist

Claude Code From Scratch 想深入了解原理？姊妹项目 How Claude Code Works

从 Julia 到 Rust：AI 智能体时代面向科学计算的可微张量栈

給 Agent 開發者的駕馭工程 (9): 自建 Agent 的框架選型: 全套 Deep Agent 還是從基礎構建?

從 Code Act 到 Claude Code Dynamic Workflows 深度技術解析

MCP Code Mode: Context Engineering for Efficient Tool Execution in LLM Agents

Programming

AI Tools Accelerates Coding, But Not Overall Software Delivery, GitLab Research Finds

Build generative UI for AI agents on Amazon Bedrock AgentCore with the AG-UI protocol

AI System Design Guide

Introducing Dynamic Subagents in Deep Agents

GitHub Copilot SDK Deep Dive: Session Memory

Modern Web Guidance

pi-automode: Bringing Claude’s Auto Mode to Pi

How we found a bug in the hyper HTTP library cloudflare

Building Persistent Page Transitions with WebGPU and Vanilla JavaScript

A fast, offline hard-or-easy call on every prompt — scored deterministically, with no model call.

Prompt Compression and Cache Tuning: Cut Your LLM API Costs by 60%

headroom: The Token-Compression CLI That Cuts Your LLM API Costs by 60–95%

HackerRank open sourced its ATS. My resume scored 90/100. Oh wait 74/100. No — 88/100. Actually 83/100.

A return to two-pizza culture allthingsdistributed

轻量级 ADR 与架构咨询例会如何支撑架构决策工作

FastAPI 之父的 PydanticAI 是真的夯

Five tools to bolster your AI coding stack

Others

如是我问提出您关于佛法、人生或心灵方面的问题，获得基于20,090卷佛经的回答

”数学不难“ 之《线性代数不难》

Depicted are 5,000 menus from the New York Public Library’s Buttolph Collection pulled from an archive of over 25,000 menus and dated from 1880-1920. Most of the menus were painstakingly assembled by one woman.

AI 冲击组织，CEO 需要第二研发部

林檎越东瀛：苹果的日本之旅

写信的生意：近代中国城市街头的代笔人

郭建龙聊中亚：每个中国人的血脉里，都藏着一个策马大漠的梦｜上海书评播客Vol.16

从昆明到腾冲，春城与热海

美国250年 (上) 一个理想与现实的国家实验

AI时代的大学课程

Sat, 04 Jul 2026 02:02:03 GMT

AI 快速迭代，大学课程也在持续更新。以下罗列著名大学公开的 AI/ML 相关课程，所有链接均为课程官网，可免费访问课件、视频或作业。

最近半年，我发现一个很有意思的现象：

大学课程更新的速度，已经开始跟不上软件工程本身的变化。

真正值得关注的，不再是「哪些大学开了 AI 软件工程课程」，而是整个软件工程学术界正在重新定义 Software Engineering。2026 年已经出现了几篇我认为未来会成为里程碑的论文。

1. SE 3.0：Towards AI-Native Software Engineering（目前最重要）

这是我认为目前影响力最大的愿景论文。

Towards AI-Native Software Engineering (SE 3.0): A Vision and a Challenge Roadmap (DOI)

作者提出了一个很清晰的演进路线：

flowchart LR
SE1["SE 1.0<br/>Human writes code"]
SE2["SE 2.0<br/>Copilot assists coding"]
SE3["SE 3.0<br/>AI teammate"]

SE1 --> SE2 --> SE3

他们认为：

SE 2.0（Copilot）的问题是：

AI 只是 Autocomplete++
人仍然管理所有复杂性
Cognitive Load（认知负担）反而增加

因此提出 SE 3.0：

Intent-centric Development
Conversation-oriented Development
AI Teammate
Compiler.next
IDE.next
Runtime.next

这已经不是 Coding Assistant。

而是：

Developer 与 AI Teammate 共同开发系统。 (DOI)

2. ASE-26：第一份完整的 Agentic Software Engineering 教材

第二篇让我很惊讶。

ASE-26: a curriculum for agentic software engineering as a discipline (arXiv)

它不是论文介绍一个想法。

而是真的设计了一整套本科课程。

论文提出：

软件工程师未来工作的主要内容已经逐渐从写代码转变为指导 Agent 工作。(arXiv)

它设计了 21 个 Module。

其中大量内容已经不是传统软件工程：

Intent
Agent Collaboration
Evaluation
Verification
Multi-agent
Evolutionary Spiral
Human Supervision

而 Coding 已经只是其中很小一部分。

如果以后真有一本《AI Native Software Engineering》，我猜它会大量参考这套课程。

3. AI-Native Software Engineering（系统综述）

另一篇值得读的是：

The Rise of AI-Native Software Engineering: Implications for Practice, Education, and the Future Workforce (arXiv)

作者分析了 2016–2026 的 48 篇论文。

最后提出：

未来的软件工程能力可以概括成九类：

Intent

↓

Specification

↓

Agent Orchestration

↓

Verification

↓

Critical Evaluation

↓

Collaboration

↓

Metacognition

↓

Continuous Learning

↓

Ethics

最重要的一句话是：

教育重点应该从 Code Production 转向 Judgment、Verification、Orchestration。 (arXiv)

这和 Aalto 的课程思想几乎一致。

4. ICSE 2026 已经开始讨论课程如何改

ICSE 2026 的 SEET（Software Engineering Education and Training）专门出现了一篇：

AI-Driven Software Development: A New Course Concept and Assessment Model (Conf.researchr)

重点不是课程内容。

而是：

考试怎么办？

以前：

学生

↓

写代码

↓

老师评分

以后：

学生

↓

AI

↓

共同开发

↓

老师评估：

为什么这样做？

Prompt 如何设计？

为什么相信 AI？

如何验证？

这说明教育界已经开始把**验证能力（Verification）**作为核心考核对象。(Conf.researchr)

5. 软件工程教育本身正在重写

还有一篇偏教育学：

A Systemic View of a Software Engineering Education Curriculum (Sage Journals)

它认为未来的软件工程课程需要加入：

Systems Thinking
Ethics
AI
Enterprise Architecture
Human Collaboration
Lifelong Learning

而不是：

更多语言
更多 Framework

一个共同趋势

把这些论文放在一起，会发现它们其实在描述同一个未来。

flowchart TD

A["Requirements / Intent"]
B["Context Engineering"]
C["Agent Planning"]
D["Multi-Agent"]
E["Tool Calling"]
F["Verification"]
G["Evaluation"]
H["Observability"]
I["Deployment"]

A --> B
B --> C
C --> D
D --> E
E --> F
F --> G
G --> H
H --> I

这里几乎已经没有「写代码」这个独立阶段了。

代码变成整个流程中的一个产物，而不是中心。

如果我是大学教授，我会这样设计教材

结合这些研究，我觉得未来一本真正经典的教材，大概率会采用下面这样的结构：

Part	内容
Part 1	Intent Engineering（需求表达、问题建模）
Part 2	Context Engineering（上下文构建）
Part 3	AI Collaboration（人与 Agent 协作）
Part 4	Agentic Software Engineering（单 Agent、多 Agent）
Part 5	Verification（测试、评审、形式化验证、AI 审核）
Part 6	Evaluation（离线评估、在线评估、A/B、Evals）
Part 7	AI Architecture（RAG、Memory、Tool、Workflow）
Part 8	AI DevOps（Observability、Guardrails、Cost、Security）
Part 9	Human Governance（责任、伦理、团队协作）

你会发现，它与经典教材如《Software Engineering》（Sommerville）最大的区别，不是增加了一个“AI 章节”，而是整个软件生命周期都围绕 AI 重构。

结合你的背景，我还有一个建议。 我认为完全可以自己整理一套 《AI Native Software Engineering Reading List》，不局限于课程，而是按主题（Intent Engineering、Context Engineering、Verification、Evaluation、Agent Architecture、Observability、AI DevOps）持续跟踪 ICSE、FSE、TOSEM、SEET、Anthropic、OpenAI、Google DeepMind 等来源。相比等待大学教材出版，这样更有机会走在这一领域的前沿。

Coursera

Best of Machine Learning & AI

deeplearning.ai

Machine Learning Specialization

Deep Learning Specialization

Stanford

CS229: Machine Learning Andrew Ng 的经典入门课，覆盖监督学习、无监督学习、强化学习。

CS231N: Deep Learning for Computer Vision Fei-Fei Li 等开设，从图像分类到目标检测、生成模型、Embodied AI。

CS224N: Natural Language Processing with Deep Learning Christopher Manning 等开设，从 word vectors 到 transformers、LLM agents、red-teaming。

CS25: Transformers United 聚焦 Transformer 架构的前沿研究，邀请领域研究者做讲座。

CS336: Language Modeling from Scratch 从零实现语言模型，覆盖数据、架构、训练、对齐全流程。

CS146S: The Modern Software Developer 面向现代软件开发的实践课程。

CS 329S: Machine Learning Systems Design Stanford, Winter 2022

MIT

6.S191: Introduction to Deep Learning MIT 官方深度学习入门，覆盖 CNN、RNN、生成模型、强化学习、LLM，配有 TensorFlow/PyTorch labs。

Berkeley

CS294/194-280: Advanced Large Language Model Agents Spring 2025，聚焦 LLM Agent 的高级课题。

Carnegie Mellon University

11-768: AI Agents Fall 2026，CMU 的 AI Agents 课程。

复旦大学

人工智能的软件基础 2026 年春季学期，从工程实践角度讲 AI 应用的软件基础。

李宏毅

Machine Learning 2026 Spring

Introduction to GenAI and ML 2025 Fall

Software Engineering with Large Language Models

Coding 已经不是课程核心。真正重要的是：

Context Engineering Evaluation Verification Agent Design Cost Safety Observability

这些才是 AI 软件工程的新基础设施。

pi-mono pi coding agent 研究

Sun, 28 Jun 2026 10:00:00 GMT

pi-mono pi coding agent 研究

https://github.com/earendil-works/pi

https://pi.dev/packages

pi-mono-agent *

https://github.com/qualisero/awesome-pi-agent

https://github.com/badlogic/pi-skills

https://github.com/Dicklesworthstone/pi_agent_rust **

https://buildwithpi.com/

https://github.com/nicobailon/pi-messenger Multi-agent communication extension for pi coding agent

https://github.com/disler/pi-vs-claude-code Comparison between open source PI agent and closed source Claude Code agent 还有一些extension

https://github.com/can1357/oh-my-pi AI Coding agent for the terminal — hash-anchored edits, optimized tool harness, LSP, Python, browser, subagents

https://github.com/jayminwest/overstory Multi-agent orchestration for AI coding agents — pluggable runtime adapters for Claude Code, Pi, and more

https://github.com/tmustier/pi-for-excel Experimental Excel sidebar agent add-in

https://github.com/nicobailon/pi-mcp-adapter

https://github.com/nicobailon/pi-web-access

https://github.com/hjanuschka/shitty-extensions

memory-mode.ts - Save instructions to AGENTS.md
plan-mode.ts - Read-only exploration mode

https://github.com/nicobailon/pi-interactive-shell

https://github.com/mitsuhiko/agent-stuff

https://github.com/tmustier/pi-extensions

https://instagit.com/badlogic/pi-mono/how-to-create-custom-extensions-for-pi-coding-agent/

https://github.com/joelhans/bmo-agent/ A self-improving AI coding agent that runs in your terminal. bmo uses LLM-powered tool execution to complete tasks, and autonomously builds new tools and skills when it encounters limitations. It features multi-provider LLM routing, sandboxed tool execution, session persistence with cost tracking, and a self-improvement loop driven by reflections and periodic maintenance.

https://github.com/nicobailon/pi-interview-tool Interactive form tool for pi-agent to gather user responses with keyboard navigation, themes, and image attachments

https://github.com/yukukotani/pi-voice Headless voice interface for the Pi Coding Agent

https://github.com/nicobailon/pi-rewind-hook

https://github.com/mikeyobrien/rho An AI agent that stays running, remembers across sessions, and checks in on its own. macOS, Linux, Android. Built on Pi.

https://github.com/tmustier/pi-agent-teams An experimental Pi extension that brings Claude Code agent teams to Pi. Spawn teammates, share a task list, and coordinate work across multiple Pi sessions.

https://github.com/jademind/pi-statusbar A macOS status bar application for Pi with a local daemon for process discovery, telemetry aggregation, and one-click session jump/focus.

https://github.com/nicobailon/pi-review-loop

https://github.com/marcusschiesser/edge-pi *** Edge-Pi is a lightweight, Vercel AI SDK based coding agent library.

https://github.com/telagod/oh-pi

https://github.com/svkozak/pi-acp pi-acp communicates ACP JSON-RPC 2.0 over stdio to an ACP client (e.g. Zed editor) and spawns pi —mode rpc, bridging requests/events between the two.

https://github.com/dungle-scrubs/tallow Tallow is a terminal coding agent that starts minimal and scales up. Install only the extensions, themes, and agents your project needs, or enable everything. It drops into existing Claude Code projects via .claude/ bridging, so nothing breaks when you switch. Ships with 51 extensions, 34 themes, and 10 specialized agents.

https://github.com/bradAGI/awesome-cli-coding-agents

https://github.com/arosstale/pi-builder

https://github.com/nicobailon/pi-subagents Pi extension for async subagent delegation with truncation, artifacts, and session sharing

https://github.com/dannote/dot-pi

https://github.com/joelreymont/pz Minimal pi coding-agent re-implementation in Zig

https://github.com/philipp-spiess/modern

https://github.com/pi0/pi-vscode Minimal VS Code extension for PI Coding Agent.

https://github.com/messense/pi-parallel-agents

// pi-parallel-agents/parallel-agents.ts
pi.registerTool({
    name: "parallel",
    label: "Parallel Agents",
    description: [
      "Run agents in parallel with different models.",

https://github.com/jo-inc/pi-reflect Self-improving behavioral files for pi coding agents. Analyzes session transcripts for correction patterns and makes surgical edits to prevent recurrence.

	pi.on("session_start", async (_event, ctx) => {
		modelRegistryRef = ctx.modelRegistry;
	});
	pi.registerCommand("reflect", {
		description: "Reflect on recent sessions and improve a behavioral markdown file",
		handler: async (args, ctx) => {
			modelRegistryRef = ctx.modelRegistry;
			const targetPath = args?.trim();

https://github.com/nicobailon/pi-model-switch A Pi coding agent extension that gives the agent the ability to list, search, and switch models on its own. With this extension, you can tell the agent things like “switch to a cheaper model” or “use Claude for this task” and it will handle the model change itself, without you needing to use /model or keyboard shortcuts.

https://github.com/vamsi/pi Port of pi-mono to Python; AI agent toolkit: coding agent CLI, unified LLM API, TUI & web UI libraries, Slack bot, vLLM pod.

https://github.com/offline-ant/pi-tmux

https://github.com/butttons/pi-kit session-recall / brief

  pi.registerTool({
    name: "recall_session",
    label: "Recall Session",
    description:
      "Load a past session's conversation by filename. Use after /recall provides the session index.",
    parameters: Type.Object({
      sessionFile: Type.String({
        description: "The session .jsonl filename from the index",
      }),

https://github.com/jo-inc/pi-mem Pi coding agent extension for plain-Markdown memory system

https://github.com/nicobailon/pi-foreground-chains A skill for Pi coding agent that orchestrates multi-agent workflows with full user visibility. Each step runs in an observable overlay where the user watches and can intervene.

https://github.com/coctostan/pi-superteam Pi coding agent extension: TDD/ATDD enforcement, specialized agent dispatch, iterative review cycles, context-aware rules

https://github.com/ghoseb/pi-askuserquestion

https://github.com/kcosr/pi-extensions

https://github.com/nicobailon/pi-skill-palette

https://github.com/default-anton/pi-librarian GitHub research subagent package for pi coding agent

https://github.com/mjakl/pi-subagent A lightweight subagent extension for the pi coding agent

https://github.com/cv/pi-action GitHub Action to invoke the PI coding agent on issues and PRs via comment triggers

https://github.com/PSPDFKit-labs/pi-skills

https://github.com/l3wi/agents-workflow A multi-agent orchestration system for Pi Coding Agent that enables parallel feature development through coordinated worktrees and spawned worker agents.

/skill:plan - Unified planning: PRD → Spec → Tasks
/skill:feature - Implement with TDD in worktree
/skill:review-codebase - Periodic quality checks and optimization opportunities

https://github.com/Huvinesh-Rajendran-12/pi-agent-core A stateful LLM agent framework for Python with tool execution, event streaming, steering/follow-up message queuing, and proxy transport. 不是pi-mono的port，参考了思想。

https://github.com/rytswd/pi-agent-extensions

https://github.com/hjanuschka/pi-qmd QMD (Quick Markdown Search) extension for pi coding agent - on-device knowledge base search with BM25, vector search, and LLM reranking. Note: QMD requires Bun runtime

https://github.com/tomsej/pi-ext

https://github.com/jademind/pi-telemetry **

https://github.com/HazAT/pi-config

https://github.com/kngzzz/pi-programmatic-agents

https://github.com/Graffioh/pi-super-curl

https://github.com/ferologics/pi-extensions

https://github.com/dnouri/ai-config

https://github.com/coctostan/pi-agent-browser

https://github.com/Cygnusfear/pi-stuff

https://github.com/zach-source/pi-agent-extensions

https://github.com/default-anton/pi-subdir-context

https://github.com/sudosubin/pi-frontier

https://github.com/netandreus/pi-cursor-provider

https://github.com/aniketmaurya/pi-agent Python port of Pi.dev ***

https://github.com/hintjen/pi-extensions Pi coding agent extensions — /prune, /explore, and more

https://github.com/openxjarvis/pi-mono-python Python port of pi-mono: fully-aligned coding agent with TUI, ai, agent, and coding-agent packages ***

https://github.com/witt3rd/pipy Python ports of pi-mono packages - LLM streaming, agent framework, TUI components, and coding assistant

https://github.com/witt3rd/pipy-agent Agent framework with tool execution, built on pipy-ai. Python port inspired by Mario Zechner’s pi-mono/agent. ***

https://github.com/w-winter/dot314 Extensions, skills, prompts, and themes for Pi coding agent. There is an emphasis here on making Pi and RepoPrompt co-operate well.

https://pi-for-excel.vercel.app/

https://lazypi.org/videos.html

https://github.com/JoviDeCroock/pi-setup

https://github.com/ruizrica/agent-pi

https://github.com/tintinweb/pi-subagents

Claude Code-style autonomous sub-agents to pi.

https://github.com/jbexta/AgentPilot

https://github.com/tintinweb/pi-supervisor

https://github.com/telagod/pi-agent-colony

https://github.com/nicobailon/pi-intercom

https://github.com/xiaotonng/pikiclaw

https://github.com/aliou/pi-harness

https://github.com/pithings/pi-vscode

https://github.com/nicobailon/pi-boomerang

https://github.com/ogulcancelik/pi-extensions

pi-goal Break work into tasks, spawn parallel worker agents
pi-web-browse Browse the web via a headless browser (CDP)

https://github.com/HenryLach/taskplane

https://medium.com/@henry_49934/welcome-to-taskplane-authors-note-ad3a0278fdd3

https://github.com/rawwerks/ypi

https://github.com/stablyai/orca

https://github.com/tmdgusya/roach-pi

Clarify and Plan

https://github.com/vaslabs/pi-web-agent

https://github.com/minghinmatthewlam/pi-gui

Electron desktop shell for pi sessions.

https://github.com/HazAT/pi-interactive-subagents

https://github.com/injaneity/pi-computer-use Let Pi control your apps on MacOS & Windows

https://github.com/xibbon/PiSwift

https://ricky42613.github.io/piserini

https://newsletter.pragmaticengineer.com/p/building-pi-and-what-makes-self-modifying

https://github.com/rcarmo/piclaw

https://github.com/withastro/flue

https://lazypi.org/videos.html

https://github.com/getpaseo/paseo

https://github.com/MasuRii/pi-permission-system

https://github.com/ifiokjr/oh-pi

https://github.com/tintinweb/pi-gitnexus

https://github.com/disler/pi-agent-observability

https://github.com/jmfederico/pi-web

https://github.com/BlackBeltTechnology/pi-agent-dashboard

https://github.com/Gentleman-Programming/gentle-pi

https://github.com/open-gsd/gsd-pi/

https://github.com/cellinlab/how-pi-agent-works Pi Agent 原理与实现：从零到一实现一个 AI Agent

https://github.com/agegr/pi-web

https://github.com/czottmann/pi-automode

https://github.com/AgwaB/pi-workflow lets Pi run named, repeatable multi-step workflows: research, code review, spec conformance checks, impact review, and project-specific team routines.

https://github.com/barvhaim/pi-openwiki LangChain’s openwiki agent based on PI harness

https://github.com/buchidonggua/dg-ai-notes Pi-Agent SDK 深度教程

https://github.com/algal/pi-openai-server-compaction adds Codex-style remote compaction for OpenAI models

https://github.com/picassio/pi-code-graph/ builds and queries code knowledge graphs

Dev weekly 2026-Jun-29

Sun, 28 Jun 2026 02:02:03 GMT

AI

GLM-5.2 & GLM-5.1 & GLM-5 https://z.ai/blog/glm-5.2

Learning on the Shop floor tobi lutke

Beyond LoRA: Can you beat the most popular fine-tuning technique?

gemma-4-webgpu-kernels Gemma 4 E2B (QAT Mobile)

CS336: Language Modeling from Scratch

你不知道的 GEO：AI 可见性的原理、实践与取舍 Tw93

Using LLMs to secure source code

Complete Agentic AI Course In 10 Hours- Langchain, Langgraph, RAG,Vectorless RAG, Guardrails,Evals

Larger Context Windows Don’t Fix RAG — So I Built a System That Does

Pydantic AI v2: capable agentic loops

I Built 11 Models to Predict the 2026 World Cup. They Crown Four Different Champions.

Migrating LiteLLM to Rust - Building the Fastest and Litest AI Gateway

拿下 OpenAI Offer 后，她复盘了 57 场面试：Transformer 要会手写，LeetCode 还得刷

【RAG 進階之路-2】如何 Monitor BM25？

Programming

Astro 7.0

StyleX

HeroUI is the modern UI library for web and mobile

Sätteri A Markdown pipeline forged in Rust for the JavaScript world

砍掉 90% 冗余词元，省下 70 万美元：Netflix 开源工具狙击 AI 账单黑洞

Notes of the book System Desgin Interview - An Insider’s Guide

ValueClaw Strategy Orchestrator — managing autonomous trading strategies, monitoring market conditions, and coordinating 97 specialized skills across 17 categories through a token-efficient progressive discovery system.

Announcing TypeScript 7.0 RC

Parse PDFs for RAG Locally with Docling: Rich Tables, No Cloud Upload

Apache Iceberg Optimization: A Guide

Building a LangChain-Style AI Agent in the Browser Using Chrome’s Prompt API and Web Workers

How OpenTelemetry traces LLM Calls, Agent Reasoning, and MCP Tools

Context Specs — the full story from capitalone

Parse Scanned PDFs for RAG with EasyOCR: Free OCR Gives You Words, Not a Document

Others

赵思渊谈明清江南土地制度中的国家与市场

播客｜「广州湾」风云：从雷州半岛和湛江历史看近代中法越的跨境变迁

《生活蒙太奇》再版序·绕了一圈，回到这些格子里

学术研究-论文-论文解析

Mon, 22 Jun 2026 01:23:45 GMT

有，而且现在已经形成了一个独立赛道：

Paper Reading AI
Paper Copilot
Research Assistant
AI Literature Review

如果你的目标是：

看 arXiv
理解算法论文
看 OpenSearch / RAG / LLM 论文
快速掌握核心思想

那么下面这些工具比直接读 PDF 效率高很多。

第一梯队：最适合工程师

Explainpaper

这是目前最经典的论文解读工具。支持上传 PDF 或论文，然后：

高亮任意一句
AI 解释这句话
调整解释难度（小白→专家）
上下文感知解释
多轮追问

例如你看到：

IndexShare is a cross-layer reuse trick for DSA

直接高亮：

什么是 cross-layer reuse？
什么是 DSA？
为什么可以减少计算？

它会结合上下文解释，而不是单独解释术语。(Explainpaper)

对于刚开始读 Transformer、RAG、检索论文的人，非常友好。

SciSpace ChatPDF

学术圈使用率可能比 Explainpaper 还高。

特点：

自动总结论文
解释公式
解释图表
对话式提问
支持数亿篇论文

很多研究生直接把它当：

论文版 ChatGPT

使用。(SciSpace)

第二梯队：适合你这种架构师

我觉得你会更喜欢下面两个。

Open Paper

特点：

论文
+
引用分析
+
AI问答
+
注释

它会告诉你：

论文核心贡献
哪些结论最重要
引用了哪些经典论文
后续哪些论文引用了它

特别适合：

BM25
ColBERT
SPLADE
RAG
GraphRAG

这种技术演进路线研究。(Open Paper)

Moonlight

最近在 AI 论文圈很火。

特点：

直接打开 arXiv
不需要上传 PDF
选中即解释
解释公式
翻译论文
AI Chat

它比较像：

Cursor for Papers

体验很好。(Moonlight)

第三梯队：适合深度研究

Elicit

如果你以后做：

文献综述
Literature Review

这是神器。

它不是解释单篇论文。

而是：

给我找
Bi-Encoder vs Cross-Encoder

相关论文

然后自动生成：

Paper	方法	优点	缺点

这种综述表格。

很多博士生直接用它做初步调研。(Reddit)

第四梯队：开源派

作为程序员，我猜你可能更感兴趣。

Artifact

特点：

开源
自托管
支持 arXiv
支持 PDF
支持个人知识库

本质：

论文阅读器
+
RAG
+
AI Assistant

而且可以自己改代码。(Artifact)

PicX

比较有意思。

上传论文以后：

自动总结
+
自动生成白板图

把论文画成思维导图。(PicX)

如果是你目前研究的内容

你最近主要在研究：

BM25
Hybrid Search
OpenSearch
Cross Encoder
RAG
PDF 搜索

我建议直接采用：

第一层
Moonlight

第二层
Explainpaper

第三层
ChatGPT o系列 / GPT-5.5

工作流：

arXiv论文
     ↓

Moonlight
看摘要和图

     ↓

Explainpaper
逐段解释

     ↓

ChatGPT
让它用中文讲给我听

例如你把：

Dense Text Retrieval Based on Pretrained Language Models: A Survey

丢进去。

然后问：

假设我是软件架构师，
不要学术语言，
给我解释：

BM25
Bi-Encoder
Cross-Encoder
ColBERT

它们的关系。

通常比直接啃 50 页 PDF 快 5~10 倍。

Moonlight（体验最好）
Explainpaper（解释能力最好）
Elicit（做综述最好）
Artifact（适合自己搭建）

这四个已经覆盖了 90% 的论文阅读需求。(Explainpaper)

Dev weekly 2026-Jun-22

Sun, 21 Jun 2026 02:02:03 GMT

AI

Vision LLMs are PDF Parsers Too: Reading Charts and Diagrams for RAG

別再看 AI 預測文了: Cedric Chin 的 Sensemaking 三部曲教你怎麼真正看懂 AI

DiffusionGemma: Google’s Diffusion-Based Open Model for Faster Text Generation

When Context Collapses: Teaching Agents to Detect and Recover from Lost Memory

Agentic coding and persistent returns to expertise

Programming

SVGs and PDFs can both be interactive

visx a collection of expressive, low-level visualization primitives for React

向量已死? Grep 萬能? 不，你需要的是「策展」一組檢索工具

Bringing more agent harnesses and frameworks to Cloudflare, starting with Flue

為下一個模型而寫,別為上一個:Anthropic 三場演講的開發心法

How Does Cursor Index Your Codebase?

Quality Playbook open-source skill that turns AI coding agents into rigorous quality engineers

Introducing the Open Knowledge Format

cocoindex-code A super light-weight embedded code search engine CLI (AST based)

Coding Agent 技术全景图：Context Engineering、Subagents 与 Harness，一年范式转移全解析

DataAgents: How we turned 9 months of analysis into 10 days

pdfslick View and Interact with PDF documents in React

HiKid 练习英语口语和听力

我开源了自己的文献检索工作流：PaperSeek

WeClaw 微信 AI Agent 桥接器

Knowledge Catalog (formerly Dataplex), is an AI-powered data catalog and metadata management platform

Zvec is an open-source, in-process vector database

OpenSharing Open standard Building on the original Delta Sharing protocol, OpenSharing defines a cross-platform, vendor-neutral protocol for secure zero-copy sharing of Data and AI assets

Deep Agents 实战

Vercel Labs 开源 Zero-Native：一个基于 Zig 的跨平台原生应用框架

liteparse

codebase-memory-mcp fastest and most efficient code intelligence engine for AI coding agents

How Perplexity Built Their Search Engine

Governing AI in the Cloud: A Practical Guide for Architects

【广发金工】构建全自动的因子挖掘AI智能体

亚马逊云科技发布新一代 Amazon OpenSearch Serverless

Formal methods and the future of programming

Phistory archives versioned system prompt snapshots from agent CLIs like Claude Code, Codex