You can launch 100 coding agents. Who keeps them moving?
Six rules for running the software development lifecycle across a fleet of coding agents, November 2026.Over the last 30 days, coding agents...
Topic
#clippings
319 篇笔记
Agents Are Going Async
Imagine a chief of staff who disappears every time you hand them a big project and only comes back once it's done. You'd probably fire them ...
Enterprise Evals 101
Imagine your worst nightmare: your agent passes every demo, then a customer asks something you never tested, and it fails in public.In April...
The components of agent memory
For Claude to learn from experience, it has to decide what’s worth remembering before it knows what’s next. Here’s how “dreaming” helps it organize memory and correct mistakes.
Designing an Always-On Agent
Chatbots wait for a message. An always-on agent is single-threaded, proactive and situationally aware. Here's how we design them, and what models still lack.
Software Factory Engineer: Job Description
An emerging engineering role for the transition from AI-assisted development to L4 Autonomous Software Factories.The RoleWe are looking for ...
ThoughtWorks 技术雷达 Vol.35 摘要
ThoughtWorks 技术雷达第 35 期（2026 年 11 月）中文摘要：四大主题、117 个条目与团队行动建议。
为团队构建真正的 Agent OS：来自 Ankole 的工程实践
此为《Building a Real Agent OS for Teams: Lessons from Ankole》的中文翻译版，倘有歧义请以英文原版为准。1. 从单用户 Harness 到集体认知在独立开发者的笔记本电脑上，Claude Code、Codex 与 OpenCl...
Agent State Machines: Building Deterministic Scaffolding for Agents
If you're not using some type of agent state machine, you're completely at the mercy of the nondeterministic nature of models. Are you reall...
Concurrent agents on one file system
I think there's a lot of confusion out there about what a file system really is. There's lot of places where we can assign blame.People's fa...
The most lucrative AI business to start in 2027
The most lucrative skill has always been capturing attention.It’s obvious why.If your goal is making more money, it doesn’t matter how good ...
Critical thinking when agents do the work
Back in the day, critical thinking was taught as approaching new info with humble curiosity and a healthy dose of skepticism - take in diffe...
The era of personal software
📝 摘要 作者以自己构建的健康数据追踪应用为例，讨论 LLM 时代「个人软件（personal software）」的兴起：当构建成本足够低，软件将逐渐商品化，越来越多人会为自身需求量身定制工具。他完整记录了实现链路——Apple Health 数据经 Health Auto Export 每小时导出 JSON，通过 端点写入 $5 的 PlanetScale Postgres；数据库 sche…
Program Status Protocol
A terminal escape sequence that lets a program tell the terminal what it is doing: idle, working, waiting on the user, finished, or failed, and why.
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
十分钟精读 DSH Super Agent：从 Paper 到 PRD
1. Super-agent作者认为，以 Coding Agent 为代表的单任务 Agent 已逐渐能够独立完成复杂工作，但现有系统仍以单一 Session 为基本单位。用户并行推进多个任务时，需要自行切换会话、管理上下文；文章称一个人通常只能同时跟进 3～5 个活跃 Sess...
Exploring Super-agent Design: Towards a More Autonomous Agent System
The original article is in Chinese. This is a machine translation.This article discusses architectural ideas. It is not an introduction to e...
New context engineering challenges in late 2026
It feels like everyone is converging on the same agent architecture: an LLM, running in a loop, calling tools, running in a sandbox. Claude ...
腾讯 2026 AI Agent 企业解决方案及落地实践手册
腾讯、腾讯云与 WorkBuddy 联合发布的 78 页《2026 AI Agent 企业解决方案及落地实践手册》精读：七类提效场景与 WorkBuddy / ClawPro / Agent Runtime / TokenHub 四件套，三阶段跃迁与三代办公，研发运维、SaaS、内容生产、AI 社交陪伴、游戏 NPC、智能客服与生态共建的方案和案例。
bliki: LLM Voice
The uncanny valley of LLMs' prose
Building an AI-native engineering org
Last week Will Larson (@Lethain) asked how leaders are evaluating engineering managers for AI skills. My reply, a list of questions to think...
Building fine-grained version control for agents with Rust and Durable Objects
How DeltaDB streams every edit from humans and agents into one causal graph.
vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model
vLLM Semantic Router released d3-ultra, a 26.09B Apache-2.0 multimodal decision model scoring 87.2 on its own SR-Text benchmark.
コードを読まずにマージするために必要なこと
📝 摘要 本文讨论「no look merge」——开发者完全不读代码就合并 AI 产出的 PR——究竟需要哪些前提。作者认为，仅把质量确认自动化（测试验证行为 + AI 做设计与安全评审、修正后重新验证、存在未执行验证或未解决指摘就不合并）并不足以移除人工审批，真正的障碍在于组织共识：如果漏检后开发负责人仍追究个人责任，那实际上等同于要求人继续读代码。文章引用 2021 年的 Ship / S…
bbtree: fastest ordered map
How bbtree, a drop-in replacement for std's BTreeMap, cuts cache misses, instructions and cache coherence traffic, and how it is kept correct.
The new set of problems for engineers to solve
A lot of software engineers are going through what can only be described as an identity crisis. The job has changed a lot. Some are depresse...
多 Agent 协作终极解：实测 YC 开源新作 Agent Relay，让 Claude Code 和 Codex 直接对话
这周，美国创业孵化器 Y Combinator 2026 年秋季营（YC Fall 2026）的创业公司 Agent Workforce 发布了开源项目 Agent Relay。公司的两位联合创始人分别是 CEO Will Washburn 和 CTO Khaliq Gant，其...
How I work with Agents
... as of October 2026.Use the latest and greatest models. Everything else is potentially a waste of time. If a model is wrong 10% of the ti...
Kubernetes GPU Guide 2026: Scheduling, Sharing, Monitoring
Run GPUs on Kubernetes in 2026: install the NVIDIA GPU Operator, share GPUs with MIG or time-slicing, serve LLMs with vLLM, and monitor with DCGM.
What FDE means at OpenAI
What forward deployed engineering means at OpenAI
用 Claude 做动画，全网最全攻略：4 条路线、11 种风格、30+ 工具和案例、全部素材来源（💗建议收藏⭐）
一条 3 分钟的梵高动画：名画在金画框里一笔一笔画出来，旋涡会动，信封会飞，拍卖槌会落下。没有下载一张图，没有打开剪辑软件，我自己也没写一行代码。画面全是 Claude 写程序“画”出来的。做完这条，我把市面上用 Claude 做动画的玩法全部扒了一遍：官方工具、开源项目、海外爆...
Agent Compute Engine deep dive | Google Cloud Blog
Traditional compute primitives are not built to run hundreds of millions of these agents at planet scale, which led us to build Agent Compute Engine.
Building agents: When to fold subagents into your main agent loop
When to consider using fewer subagents and how to migrate the into your main loop.
Lemo-Opuscar 史诗级更新：8种新风格+新改编玩法，我把风格库变成了风格制作工具箱
Hello 大家好，我是lemomo。这次 Lemo-Opuscar 迎来了上线以来最大的一次更新，不光有新的风格，还有新的玩法，整个项目定位从【视频风格库】变成【搭建自己风格库的工具】。新增【剪辑与纪实】8 种风格，目前从最早的39种已经更新到了59种。未来会继续更新，我的目标...
LLMs are making formal software verification feasible at a much larger scale
Claude formally proved security properties in the compiled code of 123 open-source libraries. Along the way, it found bugs that had eluded detection for years.
How OpenAI ingests petabytes of logs every day
In this blog we will see where OpenAI's logs go, and how one line of code saved its logging system. In June 2025, ClickHouse wrote about OpenAI's logs.
How To Use Decision Models for Evals
Co-authored with Antaripa Saha.Decision models have exploded onto the scene recently, starting with TypeSafe’s Jev, followed by Perplexity’s...
Inspect AI 调研（索引）
调研笔记已归档到知识库： 👉 20-Areas/AI/Agentic/Harness/Inspect AI 调研 一句话结论：Inspect AI 是 UK AISI 出品的 LLM 评测框架（MIT，0.3.282），核心抽象是 ，把工具调用、Agent、沙箱、日志回看做成一等公民，配 的 200+ 预置 benchmark 可直接横向扫模型。
How Uber connects its AI agents to models and tools
In this blog we will see how Uber puts two gateways around its AI agents, and how every call carries the person who asked.
形式化验证 LLM 生成的代码（FATTW 2026, Lucas Cordeiro）
FATTW 2026 会议演讲（与 ASE 2026 联合举办）：LLM 负责猜测（不变量、契约、harness），证明引擎负责裁决（ESBMC/k-induction/Vampire）。以 310,531 个 C 程序基准、Test-Comp/SV-COMP 竞赛成绩和 AWS/NVIDIA/vLLM 生产代码漏洞为证据，论证『模型建议 + 证明引擎确认』的设计原则。
AI Agents Are Having Their Kubernetes Moment
Ten years ago, we saw infrastructure responsibilities moving out of microservices and into Kubernetes, service meshes, and proxies. Applicat...
手脑分离与边界治理：企业级 Agent 平台的总体架构设计
在本地终端环境中，开发者早已习惯使用 Claude Code、Cursor 或各类命令行 Agent 进行自主编码与系统运维。模型推理循环与操作系统的 Shell、文件系统和环境变量运行在同一台宿主机上，通过本地进程树与内存字典直接驱动。这种“脑与手同机”的拓扑，在个人开发机上极...
A deep dive into OpenSEO
Explore OpenSEO's architecture: SEO data APIs, Cloudflare Workers, its audit crawler, MCP tools, AI agents, storage, and self-hosting trade-offs.
How the Claude Code team does code review
How do you review code at the volume agents produce it? On the Claude Code team, once we had high coverage and deep reviews, we started optimizing for the minimum amount of human review time that doesn't raise reverts or incidents.
How to be more ambitious: A tutorial for humans with models
I work on Claude. It's an interesting experience: Every day I learn about yet another feat accomplished with or by Claude that I hadn't considered, didn&
Redesigning source control at Anthropic
As our merges per day grew more than tenfold, source control became our biggest development bottleneck. Here, we share how we migrated to an in-house version control system in five months.
为什么我们需要 Agent 评测
Agent 评测已经成为了当前 Agent应用的基石。如果说年初大家都在关注 Harness，那目前的热点和深水区，一定是 Agent评测和 Agent Infra。我所了解到的是，基本每个做 Agent 产品的组都很看重评测的方法论以及效果。因此我专门开了一个系列，给大家详细介...
「万字总结」- 我偷来了一份成为 Agent 工程师的学习线路
我之前面试了一家做 AI + web3 业内比较厉害的公司，但是没有通过吧，面试之后面试官给我了一份《Agent工程师入门指南》。我把它整理了一下，分享出来给大家提供参考，这是一份很详细很详细的、关于如何成为 Agent 工程师的文档，并且补充上我自己的学习心得当时给我的答复是 ...
来自 every 的自我批评，不应该让每个员工一个 Agent
Every 是一家做 AI 应用和内容的公司，最近把自己用了半年的企业 Agent 公开，同时批评自己之前的做法：受 OpenClaw 的影响，Every 为员工每人一个 agent，但在实践过程中，发现了在「维护」层面产生了诸多的问题，导致这种方法不可持续。它的联合创始人兼 C...
入职大厂两个月，我的 AI 开发心得
背景起因是和组里老板进行了几次 one-on-one。他对我目前使用 AI 工具、搭建个人 Harness，以及 AI Coding 的工作流很感兴趣。主要也是因为，我现在的需求做得又快又好，已经是一些项目的主 O 了。因此，我也借这个机会，整理了一下自己的日常 AI 工作流。目...
How to Cut Your Agent Bill by 90% With Jev Engineering and Kimi K3
How to find the decisions hiding inside your agents, write questions a system one model answers well, route on calibrated confidence and han...
The 5 levels of AI agents. From a single prompt to a production agent (Complete course)
There are five words people throw around about agents right now. Context engineering, loop engineering, Jev engineering, harness engineering...
Proxmox VE — Windows 7 VM 安装指南
📝 摘要 基于 Proxmox VE 8.4.0（i440fx/SeaBIOS）部署 Windows 7 Professional SP1 x64 虚拟机的完整实录，重点解决 VirtIO 驱动安装的兼容性陷阱。核心策略是先用 IDE 模式装系统、再切换 VirtIO Block 控制器，并写入 CriticalDeviceDatabase 注册表条目以避免 蓝屏。 核心概念： Windows…
The Evaporation of Software Engineering (and the Rise of the Agentic Builder)
If you’re like me, you likely thought software engineering was a safe and secure job …In fact, I never thought I’d have to worry about a job...
You need to be delusional if you want to succeed
Tell me if any of these sound like you:1) You feel like you’ve been lied to about what a good life looks like.You’re pursuing a goal that so...
Don't Start With Multi-Agent
When one agent should become many, when it shouldn't, and the supervisor pattern that makes the split work. (8 min)
Your Agent Harness Should Repair Itself
Everything you need to understand how an agent harness can learn from production failures and repair itself. It covers procedural memory, tr...
构建你的 AI FDE 能力入门篇： 0x01 Palantir 架构中心·中文全解
本系列编译自 Palantir Architecture Center 官方文档，共 7 篇，从整体架构到各核心层逐层拆解 Palantir 的技术体系。建议阅读顺序总览：Palantir 架构中心导读 —— 全局地图，先读这篇 AIP、Foundry 与 Apollo：一体化平...
An Accidental Blackboard
Notes from my Thoughtworks colleagues on AI-assisted software delivery
How to become a Robotics Engineer in 6 months (RESOURCES)
Robotics is the least crowded high-value skill in tech right nowThe problem is that almost nobody knows where to start, because the field lo...
Building the missing layer between AI spend and ROI
Author: René SultanAI telemetry tells us which models our agents use and what they cost, but not what we are paying them to do. We built a s...
From rg to zg: Local Search Beyond Keywords
Summary: The information humans and agents need is often scattered across large numbers of local files, making it difficult to locate accura...
The Harness Playbook
Model the agentic harness like a game engine and most of its hard problems — durable state, tools, configuration, the TUI, inference — collapse into ones other people already solved.
Your Inference Server is Secretly a Learner: Open-Sourcing Reef for Continual Self-Improving Agents
1. Before all the “RSI” hype becomes realBefore today's excitement around RSI is fully realized, we want to open-source Reef, an infrastruct...
Memory Is a Compiler
tl;dr: Memory preserves the past in a form meant to be useful in the future, then recompiles what that past means as new evidence and new qu...
There's no reason for software to be slow anymore
📝 摘要 Dan Luu 论证 AI 编码智能体（coding agents）已将软件性能优化的成本降低了多个数量级——过去需要专业性能工程师数天/数周完成的工作，现在普通人输入几句话即可完成。文章通过多个案例展示这一趋势：FRE 正则引擎的 AOT 编译器切换为 ripgrep 带来 7% 提速；Azul 游戏 AI 借助 LLM 实现多线程和 10-20 项手工难以完成的优化，远超人工编写的…
The Principles of Learning Faster.
Every day, you're learning something. From books. Conversations. Podcasts. Mistakes. Articles. Videos. The people around you.But consuming m...
Agentic Kernels in Production
TL;DR: We’ve built an agentic kernel development framework that identifies model-level optimization opportunities, generates improved kernel...
Running a Software Factory Efficiently at Uber Scale
Post author: @udaykiranIntroductionAI tools are now embedded in every phase of software development at Uber. More than 70% of pull requests ...
AI Agent 沙箱怎么选？我们把 runc、vArmor 和 Kata 跑了 30 轮
在 104 线程 Linux 物理机上实测 runc、runc + vArmor 与 Kata QEMU：Kata 用近一倍启动时间和数百 MiB 内存换独立内核，vArmor 则以接近 runc 的成本收紧行为边界。
GPU 共享怎么选：KAI vs HAMi vs MIG
Kubernetes 上的 GPU 共享：KAI、HAMi 与 MIG 对比。
KV Cache Engineering for LLM Serving, clearly explained
Everything you need to understand why the KV cache grows, the 12 ways models and serving engines reduce it, what each technique actually sav...
We are in the middle of the digital renaissance, please take advantage of it
There will be 3 types of people in the future.And the divide between them is growing larger every day.If you want to win in the next 5-10 ye...
AI agents meant to replace Meta workers made “large-scale, disruptive actions”
Report shows Meta's challenges replacing people with AI agents.
AI Chip Architectures
A look at AI Chip Architectures. NVIDIA, AMD, TPUs, Trainium, Groq, Cerebras.
Do you need a software factory?
A software factory is a repeatable loop around software work. If you're building a software factory, code good enough to ship still needs hu...
Evaluating OpenWiki with WikiBench
OpenWiki is our open source agent for generating and maintaining codebase documentation. As we improve OpenWiki, we’ve needed to measure two...
Agent 的 Sandbox 到底是怎么实现的？
从 Codex、Claude Code、Gemini CLI 到 OpenHands很多人讨论 AI Agent 安全时，关注的是：“模型会不会执行危险命令？”但真正重要的问题其实是：即使 Agent 决定执行危险命令，底层系统是否允许它成功？这就是 Sandbox（沙箱）存在的...
Building Self-Correcting Memory in OpenWiki
Memory Has a Forgetting ProblemA memory system is only as good as its ability to forget.Most discussions about agent memory focus on creatin...
How to Build a Company Brain That Gets Smarter Every Week
I am going to show you how to turn every correction, decision, and useful workflow your team gives AI into one company brain that makes ever...
How we Build Agent Environments & Tasks
TLDR: This is a practical guide on how we create synthetic agent environments and tasks. By the end, we have a two-step pipeline. The first...
io_uring：Linux 的高效异步 I/O 框架
iouring：Linux 的高效异步 I/O 框架 原文: Efficient IO with iouring — Jens Axboe, 2019 作者: Jens Axboe（Linux 块设备层与 I/O 子系统核心维护者） 说明: 本文翻译并梳理了 iouring 的设计思想、核心机制与使用方式。 📝 摘要 iouring 是 Jens Axboe 于 2019 年提出的 Linux…
Moats in the age of floods
As the labs break revenue records and absorb more capabilities, there's a growing narrative that the great flood of intelligence they bring ...
SemiAnalysis 发布 AgentX InferenceXv3 研报，Mooncake 成为关键基础设施
SemiAnalysis 发布 AgentX InferenceXv3，用大规模真实 Agentic workload 重新审视推理系统性能。随着长上下文、多轮交互和 KV Cache 复用成为常态，Agentic Inference 正从“拼算力”走向系统级优化，而 Mooncake 也正在其中扮演越来越重要的基础设施角色。
How to Cut Agent Tokens by 2.7x (using an open harness)
Everything you need to understand where your agent's tokens actually go and what to do about it. It covers what a production harness owns be...
The Audit Log Page
An audit log has two readers. Most logs are designed for the wrong one. Five theses on product audit logs.
Consistent Hashing is kinda cool
Hii in this blog I will try to cover consistent hashing in detail from top to bottom. But before getting into consistent hashing, let’s firs...
Building the last WAL on S3
Nearly two years ago, we started Archil with a simple realization: it was clear that people were spending immense effort to build new, serve...
Agent 终章（Harness 成本篇）：一次百炼账单降低 88% 实战
关于优化 token，本文或许能够提供一些经验和参考
Computers Compared
Agents get computers now, and definitions vary. A Fly.io Sprite gives each agent its own hardware-isolated Linux machine that keeps its filesystem, stays reachable, and sleeps when there is nothing to do.
Recursive Self-Improvement
Agents can already edit their tools, skills, and harness. Recursive self-improvement still needs a system that can raise the verifier without capturing it.
Slop Debt
Tech debt has rules. You cut a corner to ship on time, you write a comment that says TODO fix this, and six months later someone pays it down. It is annoying but it is bounded. A senior engineer can look at a codebase with tech debt and draw a map of where the bodies are buried.
5 design patterns for long-horizon agent harness
Any agent can look brilliant on a one-shot task. Give it a week of real work and it starts to fall apart."Long horizon" means the agent keep...
Codex as a platform: build on the open agent harness
Build Codex into the products and workflows your users already know.
Domain-Driven Design matters more when AI writes your code
The domain model and Ubiquitous Language matter more than ever when an AI agent writes your code.
Give me 20 minutes and I'll make you impossible to manipulate
The only real test of intelligence is if you get what you want out of life. — Naval RavikantA core sign of low intelligence is not being abl...
wal3: A Write-Ahead Log for Chroma, Build on Object Storage
wal3, Chroma's write-ahead log, combining a 30-year-old lock-free algorithm with S3's newest conditional writes feature.
What is a Harness?
An agent harness is a piece of software that provides an environment for an AI model to operate within.
An open source rival to Claude Managed Agents just launched
TrueFoundry's open source TrueForge challenges Claude Managed Agents with vendor-neutral model choice, claiming 50% lower agent costs and no lock-in.
Formally verified distributed locks for sandboxes
I’m wrapping up my internship at Modal, where I worked on our new, massively scalable sandbox architecture.
Fragments: August 18
fragments 18 Aug 2026
From context to dreams: architecting memory for AI agents
Have you ever felt that every conversation you have with an LLM across sessions feels like starting over from scratch? LLMs have a problem: ...
Kyverno is a platform primitive, not a security tool
Where does Kyverno live in your organization? I don’t mean which cluster! On which team’s slide deck does it show up? Whose budget line? For most companies I’ve…
The Agent Harness Should Be Open
Today, we are open sourcing TrueForge, the agent harness we use at TrueFoundry to build and run general-purpose agents. Because once agents ...
What Is Reasoning
A short summary of how reasoning traces work.
Cursor Origin review: An engineer’s perspective
A detailed review of Cursor Origin, covering Git hosting, GitHub sync, cloud agents, current limitations, and whether it’s worth switching.
Designing Loops for Production-Grade Work
What became standard practice after vibe-coding a production-grade BPE tokenizer trainer without reading a single line of code.
Git at any scale
Cursor is the best AI coding agent., Git at any scale
PaddleOCR-VL 服务问题排查与部署汇总
📝 摘要 本文汇总了 PaddleOCR-VL 服务在 vLLM 部署下的效果问题排查结论及完整部署方案。客户反馈 PaddleOCR-VL 丢失版面分析能力，根本原因是直接用 OpenAI SDK 裸调 vLLM 的 接口，只调用了 VLM 识别模型而绕过了版面分析模型 PP-DocLayoutV3。PaddleOCR-VL 的正确用法是两阶段 pipeline：版面分析 + VLM 识别。文…
Designing Agent Context by Scope and Time
When an agent performs poorly, we often blame the model.Then we enlarge the system prompt. When that becomes unwieldy, we add another agent....
Optimizing for cost and intelligence
Balance cost and intelligence on the Claude Platform, with measured results for prompt caching, effort, model choice, budgets, and multi-model strategies.
How Production LLMs Reason Better at Inference Time (as researched by Google, OpenAI, and Anthropic)
8 techniques to make an LLM reason better at inference time, covered with tradeoffs and practical notes. Every one of them is running in pro...
How to Design an Agent Harness: six decisions that turn a model into a worker you can leave alone
Your agent says the job is done and the tests never ran. It's sharp for twenty minutes, then forgets a rule you gave it at the start. You re...
Patterns and problems in multiagent systems
We ran experiments on swarms of Claude agents and found coordination failures, collusion, and sabotage. Here, we share what they mean for AI safety.
The AI Engineering Skills Map
I am delighted to present The AI Engineering Skills Map. AI allows us to build software very differently today than in 2022, and everyone wi...
The Principles of Better Decisions.
Every day you make hundreds of decisions. Some are small. What book to read. Who to follow. Which podcast to listen to. Others completely ch...
Why model routing must be in the harness
Inference spend is rising as agents are deployed more widely, take on more complex tasks, and run longer autonomously. Enterprises are start...
Practical Loop Engineering
Goals, loops, and the discipline of not delegating your judgment
"Issue tracking is dead"; How the pull request became the last chokepoint in the SDLC bottleneck
CodeRabbit's new Agentic Change Management layer reimagines code review as AI agents outpace human teams' ability to plan and prioritize work.
How Compaction Works in Pi
Why compaction is needed for large language models and how Pi implements it
TDD inside the agent loop - theater or actual value?
Notes from my Thoughtworks colleagues on AI-assisted software delivery
7 rules for self-improving agent loops every AI engineer should know
By @Saboo_Shubham_ and @secchi_eliaWhen coding agents write the code, defining what's good is the real engineering job.Coding agents can now...
The human is the loop
After a few weeks away from AI, I see my usage wasn't always productive or healthy.
How to Build a Shared AI Harness for Your Team
I am going to show you how to turn the models, agents, skills, and automations living across your team's tools and workflows into one coordi...
The company brain has a permissions problem
Every company is connecting Slack, email, docs and meetings into one brain that gives everyone access to everything the company knows.Nobody...
Your Agent Remembers Everything and Understands Nothing
Agent memory is where analytics was for years, returning what you asked for and nothing more. Then analytics started surfacing which number ...
DeepSeek 刚支持了 OpenAI Responses API，但你可能还不知道它是什么
2026 年 7 月 31 日，DeepSeek 悄悄发布了 V4-Flash-0731，附带一行更新日志：原生支持 OpenAI Responses API。消息不大，但背后藏着一条值得认真讲的故事线。一套 API 格式，是怎么从聊天框时代的产物，一步一步演化成今天 agent...
Five AI rivals just backed a shared plugin standard. Here's why it matters for developers.
OpenAI, AWS, Cursor, GitHub and Microsoft back Vercel's new Agent Plugins standard for packaging AI skills and MCP servers. Here's what it means for developers.
Profiling LLM Inference with SGLang and Torch Profiler
This blog is my attempt to learn how to profile an llm served using sglang and its in built torch profiler integration that comes out of box...
Why managed agents are the next big thing in agent building
Today we’re launching Managed Deep Agents - the easiest way to build, run and deploy production agents. We have a full release blog and docs...
Why managed agents are the next big thing in agent building
Today we’re launching Managed Deep Agents - the easiest way to build, run and deploy production agents. We have a full release blog and docs...
Why your KubeVirt VMs can't move between clusters — and how EVPN fixes it
Migrate KubeVirt VMs between clusters seamlessly with EVPN/VXLAN overlays and OpenPERouter CRDs. No network tickets required.
所有想做 FDE 的人，都该读懂 Palantir 这套方法
大模型进入企业以后，最难的一步是把一家公司的数据、规则和权力关系，变成 AI 能够理解并安全执行的系统。2026 年 5 月，OpenAI 成立了一家专门负责企业部署的新公司。它准备投入超过 40 亿美元，并计划通过收购 Tomoro，一次带入约 150 名 Forward De...
Concurrency vs. Throughput: why more parallelism can make databases slower
Increasing throughput sometimes requires reducing parallelism. A look into why this is the case for Vitess + MySQL databases
Deep Agents vs LangChain vs LangGraph
Deep Agents, LangChain, and LangGraph each offer distinct approaches to building agents. In this post, we cover the key distinctions between...
Efficient Decode Context Parallelism with vLLM for Long Context Workloads
Decode Context Parallelism (DCP) in vLLM shards KV cache across GPUs by sequence dimension, enabling 3× higher throughput on long-context agentic workloads comp
LLM Inference: Past, Now, and Future — Where Does the Value Move?
LLM Inference — Past, Now, and Future: Where Does the Value Move?For the past two years, LLM inference has been one of the most contested la...
LLM Inference: Past, Now, and Future — Where Does the Value Move?
LLM Inference — Past, Now, and Future: Where Does the Value Move?For the past two years, LLM inference has been one of the most contested la...
Why Agents Started Dreaming (And What is Dreaming)?
A long-running agent gets really worse at remembering facts over time.As the context window fills up during the conversation, and especially...
Cloudflare OS: an open platform for agents, apps, and work
Cloudflare OS is an open-source platform that lets everyone in your company build apps, automate work, and safely access internal systems, shaped around what your organization knows and how it operates.
What is WAL backpressure, and why does ClickHouse Managed Postgres need it?
How ClickHouse Managed Postgres uses WAL-aware backpressure to slow client writes, protect disk space, and let the archiver recover.
The Agent Development Lifecycle has arrived on Cloudflare
Agents can write code faster than teams can review, deploy, and maintain it. Today we’re introducing the Agent Development Lifecycle and the Cloudflare primitives that underpin it.
前OpenAI推理负责人JT与Gemini pretrain负责人Rohan: Continue learning的未来
📝 摘要 本文是红杉资本对 Core Automation 联创 Jerry Tworek（前 OpenAI o1 reasoning 负责人）和 Rohan Anil（前 Google Gemini pretrain 负责人）的访谈。两人共同阐述了 Transformer 架构的根本局限——训练停止后无法真正持续学习——以及 Core Automation 寻找"Transformer 替代者…
最佳实践：从Multi-Agent到Agent Team
2w字长文，Agent Team 的最佳实践。
Behind the scenes: How we build, test, and scale Google Agent Skills
AI agents are only as good as the instructions and context you give them. When we launched Google Agent Skills, our goal was simple: encode ...
How Stripe Built Kai on Deep Agents in 1 Week
Learn how Stripe built Kai, a company-wide AI agent on LangChain, LangGraph, and Deep Agents, reaching 5,000 users in roughly 4 weeks.
nvml-mock
nvml-mock 📝 摘要 nvml-mock 是 NVIDIA 官方的 Kubernetes Helm Chart，通过 CGo 编译的 mock 共享库，在无物理 GPU 的集群中模拟完整的多 GPU 环境，使 、device plugin、DRA driver 等 GPU 相关 K8s 组件无需真实硬件即可运行和测试。它部署为 DaemonSet + ConfigMap + RBAC，在…
YC 开源 QM：当每个员工都有 Agent，公司需要怎样的协作系统？
未来一家公司里，可能每个员工都有自己的 Agent。它们能读文件、查数据库、写代码、回复邮件，也能在后台持续运行任务。问题随之而来：这些 Agent 应该看到什么？记住什么？使用谁的账号？哪些能力可以共享？危险动作由谁批准？7 月 31 日，Y Combinator 开源的 QM...
Eval Engineering: build the gate that lets your agents merge without you (full 6-step course)
The end state is easy to describe. An agent finishes a change, opens it, and it goes in without a human reading it.Not because anyone decide...
Evaluating code review agents with ReviewBench
There are more code review agents now, and we’ve been building one internally. Code review is hard to evaluate. There aren’t many benchmarks...
A practical guide to building agents
A comprehensive guide to designing, orchestrating, and deploying AI agents—covering use cases, model selection, tool design, guardrails, and multi-agent patterns.
Codex 进阶指南：作为 Multi-Agent 编排控制平面
大多数人用 Codex 的方式，是开一个会话、交代一件事、等它做完。这种用法只动用了它很小一部分能力。Codex 真正提供的是一组完整的调度原语：发现有哪些项目和会话、创建新会话、从已有会话分叉、给别的会话派活、等它们回来、观察它们的进展、纠正跑偏的那个、把任务搬到另一台机器、最...
Agency: Secure, scalable sandboxes for agents
This is part of our series on how we're AI-pilling Sierra.In our last few posts, we talked about what happens when every employee has a sing...
Why Kubernetes is the wrong tool for agentic workloads
At some point, everyone building ai infra wonders whether they should use Kubernetes. There’s no straight answer for it because it depends o...
多模型路由进入 Agent 会话之后
模型路由在单轮问答中可以按难度分配成本，但进入 Agent 多轮会话后，还要处理上下文、工具状态、缓存和隐式状态。本文拆解会话级路由的智能与工程瓶颈，以及固定强模型、隔离子任务和有状态路由三条现实路线。
The harness is all you need (mostly)
A practical GitHub Copilot workflow for prototyping, planning, implementing, and reviewing software - without chasing every new AI tool.By @...
We rewrote our agent to run entirely in a Durable Object with Pi, Agents SDK and Code Mode
We recently finished moving the camelAI agent off of virtual machines. The agent now runs inside a Cloudflare Durable Object, its filesystem...
The Orchestrator's Tax
Use subagents to protect the orchestrator's working memory, not just to save time or run tasks in parallel.
22580: From GPT2 to Kimi3, Explained
Twenty-two thousand five hundred and eighty. That’s how many GPT-2 (2019) models fit inside KimiK3 (2026). We scaled up by a factor of 22,58...
How to remember everything you read (stop trying)
If you need to remember it, it’s not important. If it’s important, you will remember it.For some reason, it’s become an obsession to remembe...
Kimi K3 Is Here: Efficient Day-0 Support on vLLM
vLLM delivers day-0 Kimi K3 serving with hybrid KDA prefix caching, DSpark speculative decoding, production-scale disaggregation, and optimized kernels across N
Merge Conflicts Are a Process Problem, Not a Git Problem
Merge conflicts aren't caused by Git. Learn how small pull requests, frequent merges, and better workflows dramatically reduce conflicts.
Why Software Factories Fail: Turning the lights back on
This is part two of Why Software Factories FailThe talk version of this post is live on youtube: https://www.youtube.com/watch?v=Ib5GBkD555M...
Practical multi-agent orchestration in Codex
GPT-5.6 Sol gets especially interesting when it has a team to work with. Codex's new Multi-Agent V2 tools give Sol and Terra a natural way t...
The new rules of context engineering for Claude 5 models
I’ve written previously about how to best prompt the newest generation of Claude 5 models and work with them iteratively to discover what yo...
Why Software Factories Fail
or: the harness is not enoughUpdate - the talk version of this post is live on youtube: https://www.youtube.com/watch?v=Ib5GBkD555M Part 1 i...
大模型是专利期十个月的制药生意
这几天流传的 DeepSeek 梁文锋讲话里有一个有意思的数据点，正好和我最近的思考交汇，因此写下一些关于整个大模型产业的一些预测，供以后证伪修正思考。他说，DeepSeek 的 API 定价标准是：买一批设备回来，十个月收回成本。也就是十个月赚来的利润可以覆盖硬件上的投入。顺带...
CSGHub cStor 存储故障排查报告
📝 摘要 本文记录了 CSGHub 集群一次由 OpenEBS cStor 3.6.0 存储故障引发的 Redis Pod CrashLoopBackOff 排查全过程，故障波及四个 CStorPoolInstance 全部进入只读状态。排查揭示三层根因：池容量达 82% 触发 cStor 硬编码的 85% 只读保护阈值（不可配置、不可自动解除）；清理容量后卷副本 rebuild 陷入状态机死锁…
Towards Automating Eval Engineering
Today we’re releasing our Eval Engineering Skill, a skill that helps coding agents build evals using context from a repository and agent tra...
The State of Agent Wikis
In April 2026, Andrej Karpathy published a GitHub Gist describing a pattern he called the LLM Wiki.In the months since, four different teams...
State machines in 2 minutes
First it was loops. Now it's graphs. Next month it'll be something else.Here's the thing: we're constantly rediscovering decades-old softwar...
Watching Go's new garbage collector move through the heap
Go 1.26 made Green Tea the default GC. We observe its cache-friendliness with perf, visualize the heap to see how Go allocates, and poke at a sparse-page problem a non-moving collector like Go's struggles with.
Controlling Reasoning Effort in LLMs
How LLMs Learn Low-, Medium-, and High-Effort Reasoning Modes
How we built our knowledge base
Authors: @hi_im_isaac_, @learnwdaniel, @gaozenghaonote: the interactive version of full technical blog available: https://www.cerebras.ai/b...
数据研发Multi-Agent架构的Harness工程实践
📝 摘要 本文基于阿里云数据研发团队的 Multi-Agent 实践，系统阐述了 Harness Engineering（智能体脚手架工程）的核心理念与落地方法。文章指出 AI 工程经历了 Prompt Engineering → Context Engineering → Harness Engineering 三代演进，并提出核心公式 Agent = Model + Harness：模型决定…
A Swarm of Agents for Multi-Angle Analysis: Building a Team of Experts from LLMs
Not about speed. About making several agents with different viewpoints argue over one decision and reach a conclusion better than any of the...
Agent 评测：方法论与体系设计
关于评测工程的体系化总结，本文或许能给出一些参考
How to Build the Loops That Just Replaced Entire Prompt Engineering
For two years, getting more out of an AI agent meant writing a better prompt.That era is quietly ending. The best engineers in the world sto...
DSLs Enable Reliable Use of LLMs
Why Domain-Specific Languages and good abstractions make LLM code generation reliable and reviewable.
A Taxonomy of Self-evolving Agents
Self-evolving agents are becoming popular. [Hermes Agent](https://hermes-agent.nousresearch.com/docs/user-guide/features/skills) enables aut...
The zero-cost fallacy: Open source software in the agentic era
Open source software is the foundation of much of our digital infrastructure. However, is AI undermining the incentive structure? What needs to change?
Getting started with loops
There’s a lot of talk right now about "designing loops" instead of prompting your coding agent. If you spend some time on X trying to pin do...
Your KV Caching Is Broken
A practitioner's guide to KV cache management. How a modern caching architecture cuts input token costs by 90% and speeds up LLM inference b...
Designing a programmable runtime for agent orchestration
This article introduces Onyx, our VM for programmable agent orchestration. And by extension, a runtime that turns orchestration into softwar...
Career advice in the age of AI
AI models get better at anything you can write a loss function for, and school is mostly loss functions: well-defined problems graded agains...
Trust Doesn't Live in Code Review
The gate you becameAgents write code faster than anyone can read it. This is not a forecast; it is the baseline now. A single agent, in an a...
Highlights from Git 2.55
The open source Git project just released Git 2.55. Here is GitHub’s look at some of the most interesting features and changes introduced since last time.
Reshaping the economics of software: Building a future-ready core with Al/works™
Shift from AI experimentation to execution. Discover how Thoughtworks AI/works™ reshapes the economics of software delivery.
Towards org-level agent harnesses
Claude Tag has grown rapidly in Anthropic. I want to share some of the reasons for this and the behavioral shifts I’ve seen with this agent ...
Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction
📝 摘要 Google Research 发布了一种将 Multi-Token Prediction (MTP) 增层改造（retrofit）到已冻结的 Gemini Nano v3 生产模型上的方法，在不修改基座模型权重的前提下，大幅加速 Pixel 设备上的端侧 LLM 推理。该方法已在 Pixel 9 和 10 系列实机部署，使 AI Notification Summaries、Proo…
Optimizing distributed AI inference: Advanced deployment patterns
Learn about the three optimization levers for distributed AI inference: prefill/decode disaggregation, KV cache strategy, and speculative decoding.
vLLM + LMCache: A Starter Guide, No GPU Required
Get started easily: a single MacBook is all you need to develop vLLM + LMCacheFor New Contributors · Covering Frontend / L1 Eviction / L2 Storage / Observability If you ever skipped LMCache because you didn’t have a GPU on hand, this guide was written for you. LMCache’s multi-platform framework has already decoupled the GPU […]
30 Core Agentic Engineering Concepts Every Developer Should Know
There are 8 billion people on the planet.Only a fraction of developers understand how AI agents actually work.Not the demos. Not the hype.Th...
How to survive AI mass replacement (& escape wage slavery)
Unfortunately, all jobs will be gone in like the next 5 seconds.At least that’s what it feels like if you take loud people on social media a...
15 AI Agent Design Patterns Every Engineer Must Know
Every team building AI agents hits the same wall.You start with one prompt and a few tools.It works. Then requirements grow. More edge cases...
简直就是教科书级别的 AI 设计规范。
Vercel 的设计系统规范，真的是我见过的教科书级样例。今天认真学习他们的 DESIGN.md，一亮一暗两套主题，彻底把颜色、字体、间距、组件、动效、文案规则，全部描述成了一套可以执行的设计方法。感觉以后我所有产品的 design.md 都可以照这个路子来写了。具体在这里：ve...
Designing distributed AI inference: Core concepts and scaling dimensions
Learn about the five-dimensional design space in modern LLM serving, including tensor, pipeline, expert, data, and context parallelism.
Migrating LiteLLM to Rust - Building the Fastest and Litest AI Gateway
LiteLLM is moving its AI gateway to Rust: 15x throughput, 11x less memory, and sub-1ms per-request overhead. No v2, no migration, your config stays the same.
Qwen Studio
Qwen Studio offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.
The 7 Types of Agent Memory: A Technical Guide for AI Engineers
Agent memory explained: the 7 types — working, semantic, episodic, procedural, retrieval, parametric, prospective — and when to build each.
How GLM-5.2 Beat Fable 5 at Website Design
GLM 5.2 ranks 1st overall on Design Arena’s single-turn, HTML Web Design (Non-Agentic) evaluation, 5 places higher than its predecessor GLM-...
知识库最尴尬的时刻：AI 写完了，我判断不了
做了再说博客更新：013丨知识库最尴尬的时刻：AI 写完了，我判断不了。 by 伊恩，查看更多内容 👉🏻 www.ianneo.xyz/zlzs-blog昨天提到了个之前踩的坑：热衷于在 obsidian 里面塞很多很多内容，各种文章，文件，各位课程文字转录稿。但发现自己输出...
AI 101: What is Recursive Self-Improvement?
How AI systems are beginning to automate coding, experiments, evaluation, and research workflows ‒ and why Anthropic, Recursive, and Sakana ...
The Agent Loop Architecture
Everyone's asking "WTF is a loop?" Here's the question nobody's asking: what runs the loop?The AI discourse has converged on loops as a core...
You don't need ten agents. You need two tracks.
Spend enough time on the internet, and you'll see someone showing off a terminal with ten coding agents running in parallel. The implied pro...
Complete Map of Graph Technology 2026: 128 Products, 7 Categories
The complete guide to graph databases, knowledge graphs, GraphAI, and more. 7 categories, 128 graph technology products, 113 vendors
Engineering End-to-End Observability for Kubernetes Workloads
Stop drowning in Kubernetes dashboards. Learn how consistent signal architecture, not more tooling, is what actually cuts incident resolution time in half.
Cursor Origin: agent-first git hosting and GitHub alternative (2026)
Cursor announced Origin on June 17, 2026—a git hosting platform built for AI agents, not just humans. Here's what it means for agentic coding teams and why it may reshape the developer tooling stack.
Building Reliable Agentic AI Systems
AI helping pharmaceutical researchers query decades of information buried in PDF reports
knowledge-catalog/okf/SPEC.md at main
Google Cloud Knowledge Catalog Tools and Samples. Contribute to GoogleCloudPlatform/knowledge-catalog development by creating an account on GitHub.
Loop Driven Development, in practice
In Loop Driven Development I made an argument: when coding agents do the implementation, the engineering that matters moves to the feedback ...
AI Semantic Cache - Plugin
Enhance performance for AI providers by caching LLM responses semantically
Launching a million sandboxes per second in constant time
They say that in the future, everyone will have their own sandbox company for 15 minutes. Let's make that simpler, here's exactly how I buil...
Advanced Observability for Production Systems
From SLIs to LLM monitoring: 9 practices that separate mature teams from dashboard collectorsMost teams have dashboards. Few have observabil...
Loop engineering: the 14-step roadmap from prompter to loop designer.
Most developers still prompt their coding agents by hand. They type, they wait, they read the diff, they type again. 9out of 10 builders hav...
SkyPilot Sandboxes: Run Agent Code on Your Own Kubernetes, at Scale
SkyPilot runs untrusted, LLM-generated code in sandboxes on the Kubernetes clusters you already own. A single cluster sustains 50,000+ sandboxes, with multi-cluster support to go further. Individual launches in less than a second. All at up to one-tenth the cost of hosted alternatives, with your code and data never leaving your cloud.
I Replaced My Entire Growth Team With 6 AI Agents. Here Is the Full System.
The week everything changed, I was not in front of my computer.Three content pieces went out before 9am. Four leads got qualified and follow...
Building cloud agent infrastructure: what's different, and what we learned
Most agent frameworks today assume a desktop. One user, one machine, one process. The agent runs while the laptop is open, writes to a local...
Every Company's First AI Strategy Should Be a Skill Library
Watch your best people work long enough and you notice something. They have patterns.A great salesperson approaches an important call with a...
When AI builds itself
Our progress toward recursive self-improvement, and its implications.
Economy of Minds: Multi-Agent Prompt Optimization explained
Wake up babe there's a new paper from Harvard: Economy of Minds (EOM). They've made a decentralized multi-agent system where agents coordin...
Faking a Filesystem For your Agent: PostgresFS
Faking a filesystem for your agent: why PostgresFS lost to a skillCo written by @SufjanFana Working with agent harnesses, it is hard not to ...
How to Build AI Agent Swarms (Complete Guide)
This is a complete A–Z breakdown of AI Agent Swarms what they are, how to use them.Why they change everything about how you work with AI.B...
A harness for every task: dynamic workflows in Claude Code
Last week, we released dynamic workflows in Claude Code. Claude can now write its own harness on the fly, custom-built for the task at hand....
Dreaming: Better memory for a more helpful ChatGPT
ChatGPT introduces a new memory system to better remember preferences, keeping context fresh and relevant across conversations.
Every Agentic Engineering Hack I Know (June 2026)
Three months ago I posted "Every Claude Code Hack I Know." It hit 913K views. @kevinrose had asked what IDE to use, and my answer was: "No I...
How sandboxes + shared memory could unlock the next big discovery
How many breakthroughs came from carrying a solution out of one field into another?A lot, but let’s check some that are relevant to AI only:...
让你的Agent从pi上长出来
太长不看：把这个丢给你的codex/claude：https://github.com/enderzcx/pi-docs-playbook别急着写 Agent，先把 pi 文档喂给 Codex 读对一开始接了个业务，只是把大模型封装了一下，以为这样就能当业务 Agent 卖出去...
A Practical Guide to Becoming an AI-Native Engineer
If AI writing everything is the answer, then why are most engineering teams shipping more bugs, more incidents, and more technical debt than...
Dynamic Workflows: Claude Code Is Moving the Agent Plan from Chat into Executable Scripts
Claude Code shipped Dynamic Workflows a few days ago. Most people's first reaction: "finally, hundreds of subagents at once."That's half rig...
Memory Is Purpose
LLMs compressed the internet into weights. Agents need to compress work into state.LLMs are among the most successful artificial memory syst...
Session-Aware Agentic Routing: Continuity-Aware Model Selection for Long-Horizon LLM Agents
How Session-Aware Agentic Routing in vLLM Semantic Router preserves long-horizon agent continuity with session memory, safe model-switch boundaries, prefix-cach
How to build a 4-agent team, that ships a feature while you sleep (Exact Setup Inside)
Four AI agents can ship a feature while you sleep. Most people never wire them up.They fire a reviewer here, a test generator there, by hand...
Setting Up and Using the Pi Coding Agent
My experience of setting up the Pi coding agent with DeepSeek, installing packages, and using it for side projects and experiments.
多 Agent 的本质不是分工，而是注意力治理
很多 AI 产品把 Agent 设计成角色，但真正决定系统能力的，是上下文、工具、状态和环境隔离。很多多 Agent 产品，都在试图把 AI 做成一个小公司。打开产品介绍页，你很容易看到类似配置：Research Agent 找资料，Writer Agent 写内容，Review...
protocols for transactional usage of object storage
building correct applications w/o external coordination
从 305 GB 到 7.4 GB：大模型 KVCache 架构演进全景
五一本打算更新一年前写的开源引擎 KVCache 实现文档，但梳理下来发现，光是 KVCache 演进的理论部分就已足够撑起一篇独立文章。在 Agent 改变搜索习惯、重塑软件行业的今天，多轮交互带来的长上下文让 KVCache 持…
Pi Agent 配置指南：打造你的专属 AI 编程助手
在玩 pi 的过程中，感觉又回到了折腾 neovim 的时期，这种自己构建乐高玩具的过程，无论何时都深深吸引着我。什么是 Pi ？pi 又叫 pi coding agent，如果你没有听说过它，那你一定听说过 openclaw。而 pi 则是 openclaw 背后的编码智能体框...
模型量化 - vLLM 运行时量化架构
前四篇分别概述了 GPTQ、AWQ、SmoothQuant、FP8 四种量化算法。读完之后你应该理解：GPTQ 用 Hessian 逐列补偿误差，AWQ 用网格搜索做 scale 迁移，FP8 用浮点格式自身解决动态范围问题。
AI 时代到底该怎么管一个工程团队
Fiona Fung 在 Anthropic 大会上讲了 28 分钟，聊了聊 AI 时代到底该怎么管一个工程团队。她做这套幻灯片时，Anthropic 还没有推出 Routines 功能。三周后，Routines 上线了。这是一个让 Claude Code 在云端按计划自动跑任务...
本地大模型：从跑起来到跑得好
0. 太长不读很多人玩本地大模型，卡住的地方不是模型太复杂，而是不知道自己卡在哪一层。你以为是在跑一个模型，其实是在同时做五个决定：硬件、内存、Runtime、模型选择、量化。每一层都有坑，每一层都可以单独优化，但它们又互相约束。后面所有问题，基本都能归到这五层里。两条读法：如果...
Why is the sky blue?
📋 文章梳理 本文从"天空为什么是蓝色"这一问题出发，系统地讲解了光的散射原理，并将其拓展到其他天体和日常现象。 核心论点 天空的颜色取决于大气中粒子的尺寸与光波长的相对关系，不同的散射机制产生不同的颜色效果。 三大散射机制 | 类型 | 粒子大小 vs 波长 | 典型例子 | 颜色效果 | |------|------------------|----------|----------| |…
Agent 文件系统：从"喂给模型记忆"到"让模型自己翻文件"
如果你在过去一年留意过 AI agent 基础设施的动向，应该注意到越来越多的团队在讨论文件系统。Turso 的 AgentFS、Anthropic 的文件系统式 MCP 改造、Vercel 的无向量数据库知识库模板、Manus 的 context engineering——它们共同指向一个方向：当 context 成本成为瓶颈时，任何能减少它的方案都会胜出。这篇文章从三代演化（裸 context → 记忆系统 → 文件系统即 context）切入，分析四家公司的设计判断，讨论四个被忽视的盲区，最后预测分层架构的走向。
Notes from inside China's AI labs
Lessons from my trip to talk to most of the leading AI labs in China.
DuckDB Internals: Why is DuckDB Fast?
Walk through DuckDB's internals: how it skips serialization overhead, parses and optimizes SQL, and stores data in columnar row groups with zone maps.
罗福莉访谈里那几句关于 memory 的话，被几乎所有人忽略了
张小珺前几天发了一期跟罗福莉的长访谈，三个半小时。小宇宙，YouTube，B 站都有，外面已经有很多解读。大家讨论得最多的是 OpenClaw 怎么改变了她的世界观，1T 是不是入场券，组织平权怎么搞，3:1:1 的卡分配比例。听了以后，挑出来想单独写一篇的，是这期里被低估最严重...
Cat Wu 面试了几百个 PM 候选人，几乎没人答对一个问题：AI 产品经理到底应该干什么？
Cat Wu 是 Anthropic Claude Code 和 Cowork 的产品负责人，和 Boris Cherny 搭档，带着团队把产品功能的交付周期从半年压到了一天。在 Lenny's Podcast 最新一期中，Cat 聊了 Anthropic 内部的速度文化、PM ...
Obsidian + AI：8 个值得搭的自动化工作流，记得收藏！
很多人还把 Obsidian 当笔记软件用。记录想法、写日记、整理读书笔记——这些当然没问题。但它能做的事，远不止这些。我现在用 Obsidian 的方式完全不一样：资料、流程模板、历史结果全存在里面，AI 要做事就从这里读取，做完再存回来。它变成了一个调度中心——不是你在记笔记...
I Went Through Every AI Memory Tool I Could Find. There Are Two Camps.
there are 450+ repos tagged "agent-memory" on github and 460+ tagged "context-management." me and my agentic best friends went through them....
Friends Don't Let Friends Use Ollama
Ollama gained traction by being the first easy llama.cpp wrapper, then spent years dodging attribution, misleading users, and pivoting to cloud, all while riding VC money earned on someone else's engine. Here's the full history, and why the alternatives are better.
上传模型 | 传神社区文档中心
如果您需要将模型上传到CSGHub，您需要先注册一个账户，登录后可创建模型仓库并上传模型内容。CSGHub平台上的模型基于Git存储库，可为您提供版本控制、分支管理和发现共享功能。您可以向模型仓库中上传任何想要的内容，包括模型文件、配置文件和任何其他文件。
Efficient Data Processing in Python: Batch vs Streaming Pipelines Explained
Every data pipeline makes a fundamental choice before any code is written: does it process data in chunks on a schedule, or does it process data continuously as it arrives? This choice — batch versus
5.2 请求路由 · Go语言高级编程
5.2 router 请求路由 在常见的Web框架中，router是必备的组件。Go语言圈子里router也时常被称为 的multiplexer。在上一节中我们通过对Burrow代码的简单学习，已经知道如何用 标准库中内置的mux来完成简单的路由功能了。如果开发Web系统对路径中带参数没什么兴趣的话，用 标准库中的 就可以。 RESTful是几年前刮起的API设计风潮，在RESTful中除了GET…
仓库地图：五层架构速览 - Hermes Agent 源码与设计
深入解读 self-improving personal agent 的工程实现
万字深度解析：TurboQuant是什么？谷歌KV Cache压缩算法（附原理+性能数据）-腾讯云开发者社区-腾讯云
谷歌TurboQuant算法突破AI压缩极限，6倍KV Cache压缩+8倍推理加速引发行业地震。这项基于极坐标量化和JL变换的创新技术，实现精度零损失的同时大幅降低大模型内存需求，被誉"谷歌DeepSeek时刻"。存储芯片市场应声下跌，但华尔街认为效率提升将刺激AI应用爆发。
Everything Is Connected
I think the title is a bit philosophical. Isn't it?Don't worry, you will get it at the end. So, imagine a scene: you are standing in a books...
为什么 Dan Koe 的内容，会让这么多人产生共鸣
最近反复在看 Dan Koe 关于 HUMAN 3.0 的几篇文章 ( 我把译文放在了二、三条，可前往阅读）。
Scaling Managed Agents: Decoupling the brain from the hands
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
《Agentic Design Patterns》中文翻译项目 - 系统介绍 AI Agent 系统的各种设计模式
《Agentic Design Patterns》中文翻译项目 - 系统介绍 AI Agent 系统的各种设计模式
Post-training 101 | Tokens for Thoughts
A hitchhiker's guide into LLM post-training, by Han Fang and Karthik A Sankararaman
Harness engineering for coding agent users
A mental model for building trust in coding agents through feedforward guides, feedback sensors, and iterative harness engineering.
2026年，普通人如何量化交易
原文作者：@gemchange_ltd原文链接：https://x.com/gemchange_ltd/status/2028904166895112617编译、注释与改写：Mr.RC, insiders.bot (@insidersdotbot) 创始人，@0xUClub前会长...
Claude Code auto mode: a safer way to skip permissions
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
How Anthropic Built 7 Layers of Memory and a Dreaming System for Claude Code (video breakdown)
A comprehensive reverse-engineering of every memory and context management system inside Claude Code's leaked harness — from lightweight token pruning to a "dreaming" system that consolidates memories…
State of Memory in Agent Harness
Agent harnesses are where AI software actually runs. Cursor, Devin, Claude Code, Codex: these environments handle context, orchestrate tools...
Build a CLI for AI agents & humans in less than 10 mins
Every CLI built this year will be called by an agent at some point. Most aren't ready for it. The interactive prompts, colored outputs, and terminal UIs that have become the standard assumptions of CL…
两家大厂加九个博主解读Harness Engineering，它到底是什么
Anthropic 和 OpenAI 几乎同一周发了各自关于 Harness Engineering 的文章。然后整个 AI 社区炸了。有人说这是 AI 工程的第三次范式转移，有人说这不过是给旧概念换了个新名字。九个博主的观点加上两家大厂的实践，拼出了一幅有趣的全景图 为什么会出现 Harness Engineering 从 Prompt Engineering 到 Context Enginee…
Natural-Language Agent Harnesses
Linyue Pan 1 Lexiao Zou 2 Shuo Guo 1 Jingchen Ni 1 Hai-Tao Zheng 1 1 Shenzhen International Graduate School, Tsinghua University 2 Harbin Institute of Technology (Shenzhen) ply24@mails.tsinghua.edu.cn…
一个从执行到记忆的完整 Agent Harness: gstack + Compound Engineering：
零零后？ 错过了BTC，不能再错过AI 。 让 AI 帮我算命，它叫我快睡比较实际
AI Agent Memory Types Explained
How AI Agents Remember? Why AI Agents Forget? A raw language model processes each request in complete isolation. The moment a response is generated, everything is wiped. For a simple chatbot, that’s f…
Effective harnesses for long-running agents
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
I Compared gstack, Superpowers, and Compound Engineering. They Solve Three Completely Different Prob
零零后？ 错过了BTC，不能再错过AI 。 让 AI 帮我算命，它叫我快睡比较实际
AIO Sandbox 技术洞察：一个 AI Agent 沙箱的设计哲学
摘要：AIO Sandbox 是火山引擎开源的 All-in-One Agent 沙箱，将 Browser、Shell、Jupyter、VSCode、MCP 等能力打包进单个 Docker 容器。本文从源码角度分析其六大设计亮点：① 单容器多进程——用资源换掉分布式协调复杂度；② Fern 代码生成——API 定义作为唯一真相，多语言 SDK 自动生成；③ Shell/Bash 双轨——区分有状态…
The Ultimate Guide for a Working Agent Memory
构建有效Agent记忆的终极指南：分层记忆栈、Research Packets、5级晋升阶梯和质量过滤系统
TurboQuant：谷歌KV Cache压缩算法深度解析
谷歌研究院于2026年3月发布TurboQuant算法，运用极坐标量化与JL变换技术，实现大模型KV Cache六倍压缩及注意力计算八倍加速，且保持精度无损。
How Memory Works in DeerFlow?
mem0 团队深度解析 ByteDance DeerFlow 的记忆系统实现：基于 JSON 的本地存储、置信度评分、异步提取和 2000 token 注入预算
How we build evals for Deep Agents
LangChain 团队分享 Deep Agents 的评估方法论：如何设计、分类和运行 Agent 能力评估
Thoughts on slowing the fuck down
Thoughts on slowing the fuck down
From "Reasoning" Thinking to "Agentic" Thinking
The last two years reshaped how we evaluate models and what we expect from them. OpenAI's o1 showed that "thinking" could be a first-class capability, something you train for and expose to users. Deep…
How Middleware Lets You Customize Your Agent Harness
Agent harnesses are what help build an agent, they connect an LLM to its environment and let it do things. When you’re building an agent, it’s likely you’ll want build an application specific agent ha…
TurboQuant - 谷歌推出的向量量化算法
TurboQuant 是 Google Research 推出的向量量化算法，可将大模型 KV Cache 从 32-bit 压缩至 3-bit，实现内存降低 6 倍、推理速度提升 8 倍，且精度零损失。
按头学习| 让 AI 真正做好复杂任务，需要一套什么样的系统？（Anthropic官工程师实践指南）
同一个游戏，$9 做出来的核心玩法是坏的，$200 做出来的能玩。 Anthropic 工程师用实战告诉你，多智能体协作框架是怎么工作的。 同一个任务，两种做法，结果让人无法忽视。 第一次：让一个 AI 单独去做，花了20分钟，花了 9美元。游戏界面出来了，但核心玩法是坏的，角色根本不响应操作。 第二次：用一套多智能体系统去做，花了6小时，花了 200美元。游戏能玩了，画面有质感，内置了 AI 辅…
How We Built Our Sandboxed Claude Agent without Leaking Its API Keys
Listen Labs 分享如何构建安全的沙盒化 Claude Agent，通过 Workstation 架构实现 API 密钥零暴露
Demystifying evals for AI agents
Demystifying evals for AI agents
Harness design for long-running application development
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
the missing layer in your agentic stack
Creator of Kaku · Mole · Pake · MiaoYan
CLAUDE CODE 最佳实践：从"能用"到"真的好用"
一开始我以为 Claude Code 就是一个"更高级的 Copilot"——在终端里打字，它帮我写代码，完事。后来才发现，这东西的上限远比我想象得高，但前提是你得知道怎么用。 这篇文章结合官方文档和实战经验，整理了我认为最重要的使用技巧。有些是血泪教训，有些是读文档才知道的隐藏功能。 一、CLAUDE.MD：最被低估的功能 如果你只记住这篇文章的一件事，就记这个：写好你的 CLAUDE.md…
We broke the frontier in agent memory: Introducing ~99% SOTA memory system.
Supermemory 团队发布 ASMR (Agentic Search and Memory Retrieval) 技术，在 LongMemEval 基准测试达到 ~99% 准确率，无需向量数据库或嵌入
Avoid Mini-frameworks
很多小团队在工作中，往往会发明自己的"定制框架"。 他们原来使用的是通用框架，但有不满意之处，于是决定在通用框架基础上定制自己的框架。 这种"定制框架"有一些共同特点： （1）由小团队创建，旨在解决他们的痛点； （2）底层是其他更通用的技术栈或框架； （3）引入原有技术栈不存在的新概念和术语； （4）创建者声称这个定制框架"神奇地"解决了许多问题，并推广更多人使用它。 我的个人经验是，"定制框架"…
你不知道的 Agent：原理、架构与工程实践
Creator of Kaku · Mole · Pake · MiaoYan
5 Agent Skill design patterns every ADK developer should know
Old man yelling at Claudes. The fucking bluecheck is temporary...https://t.co/mnOoWUqt4ghttps://t.co/8i5vIRDt6P
使用 pi 打造属于你自己的 Coding CLI
当前主流的 AI 编程 CLI 工具，工程复杂度已经远超必要。 最基础的交互无非是发送 prompt、获取回答，但层层封装的 harness 架构把这些工具变成了吞金巨兽。发一句 hi 就消耗掉当天 10% 的模型配额；MCP 服务器空挂着不做事，常驻占用 15% 左右的上下文窗口；Plan Mode、Subagents 本质上只是上下文管理的技巧。 为了让 AI 辅助编程，你真的需要这些吗？ �…
Lessons from Building Claude Code: How We Use Skills
Creator of Kaku · Mole · Pake · MiaoYan
Harness Engineering 深度解读：AI Agent 时代的工程范式革命
\> 基于六篇核心文献的综合分析：OpenAI (Ryan Lopopolo)、Anthropic (Justin Young)、Martin Fowler (Birgitta Böckeler)、LangChain、Latent Space、Cassie Kozyrkov 一、什么是 Harness Engineering？ Harness Engineering 是 2025-2026 年 A…
智能体工程的 8 个等级【译】
作者：Bassim Eledath 原文：The 8 Levels of Agentic Engineering — Bassim Eledath AI 的编程能力正在超越我们驾驭它的能力。这就是为什么所有那些拼命刷 SWE-bench 分数的努力，并没有与工程领导层真正关心的生产力指标同步。Anthropic 团队用 10 天就上线了 Cowork，而另一个团队用着同样的模型却连一个 POC（概…
Installed Paperclip, Now What !?
My Paperclip article hit 2.7 million views. The comment section was mostly made up of "cool demo", "it doesn't work", and "how do you handle" questions... This follow up article is to get some ideas g…
从零开始两天构建一个 Claude Code：带你拆解 AI CLI 的每一层
前两天突发奇想：一个生产级的 agentic CLI 到底需要哪些组件？每一层的具体怎么实现？SSE 缓冲区怎么管理、system prompt 怎么分段、工具权限怎么拦截、上下文满了怎么压缩。这些问题靠读文档回答不了，靠逆向混淆代码效率极低。 所以选择了另一条路：以 Claude Code 为参照系，从零重建一个功能等价的实现——纯 TypeScript，零框架，唯一的依赖是 fast-glob…
你不知道的 Claude Code：架构、治理与工程实践
今天这篇文章源于最近半年深度使用 Claude Code、两个账号每月 40 刀氪金换来的一些踩坑经验，希望能给大伙一些输入。 刚开始我也把它当 ChatBot 用，后来很快发现不对劲：上下文越来越乱、工具越来越多但效果越来越差、规则越写越长却越不遵守，折腾了一段时间，研究了 Claude Code 本身之后才意识到，这不是 Prompt 问题，而是这套系统的设计就是这样的。 这篇文章想和大伙聊聊…
KV Caching in LLMs, Clearly Explained
You must have seen it every time you use ChatGPT or Claude that the first token takes noticeably longer to appear. Then the rest stream out ...
The Anatomy of an Agent Harness
TLDR: Agent = Model + Harness. Harness engineering is how we build systems around models to turn them into work engines. The model contains the intelligence and the harness makes that intelligence use…
The Decentralized self-improving AI system that builds itself Democratically
preface: what decentralization changes a self-improving system in one repo is impressive. a network of self-improving systems across many forks is qualitatively different. in the centralized model, pr…
How I use Obsidian
My personal Obsidian vault template. A bottom-up approach to note-taking and organizing things I am interested in.
Harness Engineering Is Cybernetics
Reading OpenAI's harness engineering post, I kept having a feeling I couldn't place. Then it clicked: I'd seen this pattern before. Not once — three times. The first was James Watt's centrifugal gover…
How to design Experiments to evaluate your Agentic Systems
Our forefathers and foremothers laid down the script for improving any system: (1) Build, (2) Evaluate, (3) Refine. Step 1 is trivial now thanks to coding agents. Step 3 is also simple too once you fi…
The MCP Abstraction Tax
Every layer from App to API to MCP loses fidelity. An exploration of what gets lost and why it matters for enterprise APIs.
How To Be A World-Class Agentic Engineer
IntroductionYou're a developer. You're using Claude and Codex CLI and you're wondering everyday if you're sufficiently juicing the shit out ...
Harness engineering: leveraging Codex in an agent-first world
By Ryan Lopopolo, Member of the Technical Staff
everything is a ralph loop
I’ve been thinking about how I build software is so very very different how I used to do it three years ago.No, I’m not talking about acceleration through usage of AI but instead at a more fundamental level of approach, techniques and best practices.Standard software practices
Mount Mayhem at Netflix: Scaling Containers on Modern CPUs
Mount Mayhem at Netflix: Scaling Containers on Modern CPUs Authors: Harshad Sane, Andrew Halaney Imagine this — you click play on Netflix on a Friday night and behind the scenes hundreds of …
Inside vLLM: Anatomy of a High-Throughput LLM Inference System
From paged attention, continuous batching, prefix caching, specdec, etc. to multi-GPU, multi-node dynamic serving at scale.
Editing Files at 1000 Tokens per Second
A new model and inference method for high-accuracy full-file edits at 1000 tokens/s.
Building and operating a pretty big storage system called S3
Three distinct perspectives on scale that come along with building and operating a storage system the size of S3.
I want to build an AI agent today (full course):
从零开始构建AI Agent的完整入门教程，涵盖核心原理、5种工作流模式、Anthropic/OpenAI SDK实践、工具使用、记忆设计和多Agent协作
How Warp builds self-improving agents on Claude
How Warp devised a simple development pattern that anyone can use to create self-improving agents.
The AI-Native SDLC playbook | Claude by Anthropic
How to transform your software development lifecycle with AI—stage by stage.
How Anthropic enables self-service data analytics with Claude
Tips and approaches to maximizing Claude’s ability to drive self-serve data insights