#AI
#公众号
LangSmith 追踪调试：智能编码代理失败复盘方法
很多人第一次需要 **LangSmith 追踪调试**，是从一次智能编码代理失败开始的。最后的代码差异可能是绿色的，测试也可能通过了。但你不知道中间发生了什么：父代理为什么拆出子代理，子代理读了哪个辅助函数，工具调用有没有走偏，失败测试又是怎么被“修好”的。
Read More →
🔬
Technology
AI, systems, networks, security, and engineering practice.
75 articles total
#AI
#公众号
GPT-5.6 模型分层：能力、成本与推理档位怎么配
2026 年 7 月 9 日，OpenAI 宣布 GPT-5.6 全面可用。此前它经历过有限预览，现在以 Sol、Terra、Luna 三个层级进入 ChatGPT、Codex 和 API。Sol 是旗舰模型，Terra 面向日常工作，Luna 追求速度与成本效率。
Read More →
#AI
#公众号
Google Cloud 的 BGP 路由策略，解决了混合云三类老问题
混合云网络最麻烦的时刻，常常发生在线路还通、路径已经变得难以解释的时候：某些前缀被学进来，备用链路承担了不该承担的流量，回程路径绕过了原来的防火墙。**BGP 路由策略**解决的就是这类问题。它把路由过滤、路径优先级和回程对称性，放到 Google Cloud 的 Cloud Router 里，用策略规则统一控制。
Read More →
#AI
#公众号
Claude 的脑中，有个J-space静默工作区
当一个 AI Agent 给出一段很正常的回答时，它可能已经在内部识别出“这是测试场景”，也可能短暂考虑过伪造数据、隐藏目标或绕开评估。问题在于，团队通常只能看见模型写出来的字，看不见它没说出口的判断。Anthropic 这篇研究把问题往前推进了一步：Claude 内部存在一组可以被读取、被干预、会参与推理的表征。研究者把它叫做 **J-space**。你可
Read More →
#AI
#公众号
Claude Code 实战：先找到你的未知点
很多人用 Claude Code 写代码时，已经会给需求、贴上下文、要求先出计划。问题出在下一层：Claude Code 未知点没有被提前暴露出来，长任务就会在实现中靠猜测前进。这里的未知点不是“模型不知道某个知识点”。原文把它放在“地图”和“领土”的差距里讲。地图是你给 Claude 的提示词、技能和上下文；领土是代码库、真实业务和各种实际限制。两者之间没
Read More →
#AI
#公众号
公众号文章转视频：我跑通了一套 AI 视频生产工作流
我最近把一篇 LangChain 文章，改成了一条 70 秒左右的竖屏视频。这次不是把公众号配图裁成 9:16，也不是随便加一段音乐，让它看起来像视频。而是完整跑了一遍流程：文章先变成口播稿，口播再拆成分镜，分镜生成关键图，关键图变成短视频段，最后用工程工具统一处理字幕、音频、BGM 和导出。
Read More →
#AI
#公众号
How Deep Agents Manage Long Context
Source: LangChain BlogOriginal title: Context Management for Deep AgentsURL: https://www.langchain.com/blog/context-management-for-deepagents
Read More →
#AI
#公众号
Four Claude Code loops: how to hand off checks, stop conditions, triggers, and recurring work
Source: Claude BlogOriginal link: https://claude.com/blog/getting-started-with-loopsPublished: June 30, 2026Topic: How Claude Code defines agentic loops, and how to choose between
Read More →
#AI
#公众号
Building Agentic AI Networks Starts with a Shared Platform
Source: NVIDIA Developer BlogOriginal: https://developer.nvidia.com/blog/how-telcos-build-autonomous-networks-with-agentic-ai/
Read More →
#AI
#公众号
Evaluating AI agents in real environments with Harbor
Source: LangChain BlogOriginal: https://www.langchain.com/blog/unified-stack-for-evaluating-agentsPublished: June 30, 2026
Read More →
#AI
#公众号
How Deep Agents run untrusted code
Letting an agent write a small script to coordinate several subagents is a natural next step for agent workflows. It reduces the back-and-forth of one tool call at a time. It also
Read More →
#AI
#公众号
Dynamic Subagents in Deep Agents: Code Handles Coverage, Models Handle Judgment
When agents are asked to review a directory, summarize hundreds of pages, process a batch of support tickets, or run a security scan, the visible failure is often simple: the task
Read More →
#AI
#公众号
NVIDIA's AI Agent Security Architecture for Enterprise Workspaces
When enterprises connect agents to real systems, the problem shifts from answer quality to execution control.A chatbot can give a wrong answer. An enterprise agent can read code, r
Read More →
#AI
#公众号
Loop Engineering: The 14-Step Path from Prompt Engineer to Loop Designer
Many developers already use coding agents every day, but the operating pattern is still manual: describe a task, wait for a change, read the diff, and send the next instruction. Fo
Read More →
#AI
#公众号
Agent cost optimization starts with a stable prompt prefix
A long-running agent may add only a small amount of new information on each turn.The model still receives much more than the latest message. System instructions, tool schemas, load
Read More →
#AI
#公众号
Cloud cache optimization needs a cost model, not only a hit-rate target
Google Research shows how cost-aware TTL decisions reduced Spanner memory usage by 15.5% and total cost of ownership by about 5%.
Read More →
#AI
#公众号
Human-Agent Teams Need Management Systems, Not More Chat Windows
Working with AI used to mean one person talking to one assistant. The next shift is different: multiple humans and multiple agents working in the same workspace toward shared goals
Read More →
#AI
#公众号
From Trace to Durable Context: How to Build Memory into AI Agents
Many agent projects run into the same practical problem: the agent makes the same mistake again after the user has already corrected it.
Read More →
#AI
#公众号
Giving Codex a Durable Place to Work
OpenAI's white paper "Codex-maxxing for long-running work" is less about a single coding trick and more about a working pattern: Codex becomes useful when a task has somewhere to l
Read More →
#AI
#公众号
Claude Code's New Canvas: Live Team Pages That Update
Claude Code artifacts turn an agent session into a page the team can open, inspect, and revisit.That matters because AI coding work often disappears into a chat transcript. A devel
Read More →
#AI
#公众号
Where Claude Code instructions should live
Teams often make Claude Code harder to steer by putting every instruction in one place. Build commands, release steps, file-specific rules, personal preferences, and security restr
Read More →
#Cloud
#Observability
#Networking
Google Cloud Network Insights: End-to-End Observability for Cross-Cloud Networks
Google Cloud Network Insights uses active synthetic probing and Monitoring Points to make hybrid and multicloud network paths observable from source to destination.
Read More →
#AI
#公众号
Claude Code Makes Domain Judgment More Valuable
Claude Code is making domain judgment more valuable.That is the core takeaway from Anthropic Research's report, "Agentic coding and persistent returns to expertise." The report ana
Read More →
#AI
#公众号
Claude's Visible Thinking: More Reasoning, Not Full Transparency
When Claude shows its reasoning, the tempting interpretation is simple: now we can see how the model thinks.That interpretation is too strong.
Read More →
#AI
#公众号
Anthropic's Think Tool: An Auditable Checkpoint Between Agent Tool Calls
An agent that can call tools is not automatically reliable. The hard part often appears between two tool calls: the model receives a tool result, interprets it, applies policies, a
Read More →
#AI
#公众号
From Workflow to Agent: An Engineering Choice Checklist
Many teams start an agent project by choosing a framework and wiring together tools, memory, planning, and an execution loop. The system looks complete, but the first real debuggin
Read More →
#AI
#公众号
Anthropic's Contextual Retrieval: Giving Every Chunk Its Context Back
Many RAG systems fail before the generation model starts writing. The system retrieves the wrong chunks, or it retrieves chunks that look relevant but lack the surrounding informat
Read More →
#AI
#公众号
AI agents need sandboxes before they run code
AI agents become much more useful when they can take action. One of the most valuable actions is code execution: writing a small script, running a data transformation, inspecting f
Read More →
#AI
#公众号
Benchling Shows Why Scientific Agents Need Workflows, Not Just Stronger Models
Scientific agents are hard to build with a single stronger model.Benchling's recent conversation with LangChain points to a more useful answer: production AI systems need to know w
Read More →
#AI
#公众号
Claude Managed Agents: Why Enterprise Agents Need a Runtime, Not Just a Loop
Many teams start agent work by choosing a model and wiring tools into it.That is a useful starting point. But Claude Managed Agents points to a deeper production problem: an agent
Read More →
#AI
#公众号
Headless Tools: Why Agents Need Access to the User Runtime
Most agent systems start with a familiar question: what tools should the model call?Teams connect databases, internal APIs, ticketing systems, CRMs, knowledge bases, and MCP server
Read More →
#AI
#公众号
How an Anthropic Seller Turned Claude Code into a Sales Workflow System
Many teams talk about using AI to redesign business operations. The useful starting point in this Anthropic story is much smaller: find one task that happens every day, consumes ho
Read More →
#AI
#公众号
Gemini Enterprise Agent Platform: Reliable Responses with Agentic RAG
Google Research published an experiment around Agentic RAG in Gemini Enterprise Agent Platform. The useful signal is not a generic "RAG is getting better" story. It is more specifi
Read More →
#AI
#公众号
Getting Started with Claude Cowork: Choose Tasks by Their Shape
Many teams already know how to use an AI chat window. They paste in a question, get an answer, copy the answer somewhere else, rewrite it, and then move it back into the real work
Read More →
#AI
#公众号
How to Build a Custom Agent Harness: A Practical Reading Note
Many agent prototypes start the same way: give a model a system prompt, register a few tools, and let the model call those tools in a loop until it returns a result. That can work
Read More →
#AI
#公众号
LangGraph Fault Tolerance: Retries, Timeouts, and Compensation
Production agents fail in places prototypes rarely cover: network calls, tool execution, LLM rate limits, frozen subprocesses, and external systems that only partially complete an
Read More →
#AI
#Agents
#LangChain
Introducing Rubrics: Build Agents that Evaluate and Correct Their Work
LangChain RubricMiddleware turns agent self-correction into a bounded grader loop with explicit criteria, tool-backed evidence, and per-criterion feedback.
Read More →
#AI
#公众号
Claude Code Skills: A Practical Playbook for Turning Team Knowledge Into Agent Workflows
Many teams start using coding agents and quickly run into the same problem: one-off tasks work, but experience does not compound. Every investigation, deployment, review, or verifi
Read More →
#AI
#公众号
Claude Code Dynamic Workflows: When an Agent Should Build Its Own Execution Harness
Claude Code now supports dynamic workflows: task-specific JavaScript workflows that can spawn and coordinate subagents, choose models, use isolated worktrees, verify outputs, and s
Read More →
#AI
#公众号
Production AI Agent Architecture: What Rippling Learned From Shipping Deep Agents
Rippling's production AI system is a useful case study because the hard part is not the chat interface. The hard part is running an agent across HR, IT, payroll, finance, global op
Read More →
#AI
#公众号
The hard part of enterprise agent platforms is safe self-service
Lyft's LangChain guest post is useful because it focuses on a production problem: how to let non-technical domain experts build and iterate customer support agents without removing
Read More →
#AI
#公众号
Interpreter Skills Turn Agent Workflows Into Reviewable Code Paths
Interpreter skills are an attempt to solve a common agent engineering problem: prompts can describe a procedure, but they do not guarantee the agent will run the same procedure eve
Read More →
#AI
#公众号
Claude Code Dynamic Workflows: Start With Scoped, Verifiable Work
Claude Code dynamic workflows are best understood as a way to turn a large engineering task into a temporary team of coordinated agents.
Read More →
#AI
#公众号
Using LLMs to Secure Source Code: The Bottleneck Has Moved After Discovery
LLMs can now help security teams read code, identify suspicious paths, draft proof-of-concept exploits, and suggest patches. The harder question is no longer only how to find more
Read More →
#AI
#公众号
Zero Trust for AI Agents: Study Notes from Anthropic's Deployment Framework
Anthropic's "Zero Trust for AI Agents" PDF is best read as a practical learning document, not as a short product announcement. The blog post is brief, but the PDF lays out a full f
Read More →
#AI
#公众号
How Anthropic Contains Claude Across Products
Anthropic's engineering article is useful because it treats agent safety as a systems problem, not only a model behavior problem.
Read More →
#AI
#公众号
AI Agent Engineering Series Extra: How Anthropic Built Its Multi-Agent Research System
This learning note studies Anthropic Engineering's "How we built our multi-agent research system." The article explains how Claude Research uses a lead agent, parallel subagents, m
Read More →
#AI
#公众号
AI Agent Engineering Series 05: Harness Design for Long-Running Application Development
This article studies Anthropic Engineering's "Harness design for long-running application development": how planner, generator, and evaluator agents can be arranged around a model
Read More →
#AI
#公众号
Google's AI-Era Network Infrastructure: A Learning Note
Google's article is useful because it frames AI infrastructure as a network problem, not just a compute problem.The core idea is simple: AI workloads need a network that can organi
Read More →
#AI
#公众号
AI Agent Engineering Series 04: Agent Skills for the Real World
This article studies Anthropic Engineering’s Agent Skills: a way to package repeated task knowledge, scripts, templates, references, and operating rules into reusable capabilities
Read More →
#AI
#公众号
AI Agent Engineering Series 03: Code Execution with MCP
This article studies how code execution can make MCP-based agents more efficient by keeping tool definitions and intermediate results out of model context.
Read More →
#AI
#公众号
AI Agent Engineering Series 02: How to Write Tools for Agents
This article studies Anthropic Engineering’s practical method for designing AI agent tools as contracts between deterministic systems and non-deterministic agents.
Read More →
#AI
#Agents
#Anthropic
AI Agent Engineering Series 01: Context Engineering
A full English learning note based on Anthropic Engineering's Effective context engineering for AI agents.
Read More →
#AI
#公众号
Digital Content Needs an Identity Layer
Google's latest update is easy to misread as another AI detection feature. The more important shift is deeper: digital media is starting to need an identity layer.
Read More →
#AI
#公众号
ZCube: Understanding the Network Bottleneck in LLM Inference
This is a learning-oriented rewrite of Z.ai's article on ZCube, a network architecture designed for large-scale LLM inference clusters. The core idea is simple: as inference moves
Read More →
#AI
#公众号
Claude Is Becoming an Agent Production Line
The previous piece looked at Anthropic's external map: compute, enterprise systems, delivery partners, industry workflows, and the Stainless acquisition as a connectivity move.
Read More →
#AI
#公众号
Anthropic Is Building an Enterprise AI Value Chain, and Stainless Is the Latest Piece
Anthropic's acquisition of Stainless looks like a developer tooling deal at first. It is more than that.Stainless turns API specifications into SDKs, command-line tools, and MCP se
Read More →
#AI
#公众号
What Is an Agent Harness?
LangChain's framing is simple: **Agent = Model + Harness**.The model provides intelligence. The harness turns that intelligence into usable work.
Read More →
#AI
#公众号
Databricks Shows Where Enterprise Agents Still Break
OpenAI's Databricks case study looks like a model-performance story. GPT-5.5 reached a new state of the art on OfficeQA Pro, Databricks' benchmark for complex enterprise agent task
Read More →
#AI
#公众号
Claude Computer and Browser Use: A Practical Reliability Checklist
Anthropic's guide to computer and browser use with Claude is best read as an engineering checklist. It covers what has to be true before a visual agent can reliably click, type, na
Read More →
#AI
#公众号
How to Set Up Claude Code for Large Codebases
Anthropic's article on Claude Code in large codebases is best read as an operating guide, not as a product announcement. It explains how teams make Claude Code useful in multi-mill
Read More →
#AI
#公众号
Codex Is Moving From Desktop Tool To Work Orchestrator
OpenAI's update that brings Codex into the ChatGPT mobile app looks like a mobile feature. It is more than that.The real shift is that coding agents are starting to leave the deskt
Read More →
#AI
#公众号
Anthropic Is Building A Tiered Delivery System For Claude
Anthropic's Claude for Small Business looks like a product launch. Read alongside Anthropic's May 4 announcement of a new enterprise AI services company with Blackstone, Hellman &
Read More →
#AI
#公众号
Voice Agents Are Moving From Conversation to Workflow
OpenAI's new realtime voice models are less about natural-sounding speech and more about turning voice into a working interface.
Read More →
#AI
#公众号
AI Agent Outputs Are Becoming Interfaces
Thariq's essay on using HTML with Claude Code looks, at first glance, like a file-format preference: ask Claude Code for HTML instead of Markdown.
Read More →
#AI
#公众号
OpenAI’s GPT-5.5-Cyber Is Really About Permissioning Cyber Capability
OpenAI’s announcement of GPT-5.5 with Trusted Access for Cyber and the limited preview of GPT-5.5-Cyber is not just another model release. The more important shift is access contro
Read More →
#AI
#公众号
OpenAI 发布 MRC：大模型竞争，拼到数据中心网络了
你以为大模型公司的竞争，还是谁的模型更会写代码、谁的上下文更长、谁的推理更聪明？OpenAI 这篇工程文章提醒了一件更底层的事：模型能力继续往上堆，瓶颈已经不只在算法，也不只在显卡数量，而是在数据中心网络。说得更直白一点：你买到十万张显卡，不等于你拥有十万张显卡的训练能力。
Read More →
#AI
#公众号
GPT-5.5 Instant 的真正变化：OpenAI 在争夺默认入口
OpenAI 发布 GPT-5.5 Instant，表面上看是一次默认模型升级。但这件事真正值得关注的地方，不是“又出了一个更聪明的模型”，而是 OpenAI 正在继续强化 ChatGPT 的默认入口地位。默认模型不是最炫的模型，却是最重要的模型。
Read More →
#AI
#公众号
Anthropic 把金融智能体（Agent）做成模板，企业 AI 落地正在换挡
Anthropic 这次发布的金融服务智能体（Agent），不应该只看成一次行业方案更新。更准确的判断是：企业 AI 正在从“给员工一个更强的聊天助手”，转向“把高频流程拆成可复用、可审计、可接入系统的工作模板”。这才是这次发布真正值得关注的地方。
Read More →
#AI
#公众号
OpenAI 发布 GPT-5.5：Agent 时代，模型要开始接管复杂任务
OpenAI 这次发布 GPT-5.5，表面上还是一次模型升级。但如果只看“更聪明、更会写代码、更会做研究”，就会漏掉真正重要的变化：OpenAI 正在把模型竞争，从单次回答能力，推到“长时间执行复杂任务”的系统竞争。GPT-5.5 的关键词不是 chat，而是 agentic work。
Read More →
#AI
#公众号
模型为什么开始说“哥布林”：OpenAI 暴露了后训练的隐秘风险
OpenAI 最近写了一篇很奇怪、但其实很重要的文章。主题看起来像个内部趣闻：从 GPT-5.1 开始，模型越来越喜欢在回答里提到 goblins、gremlins 这类小怪物。到 GPT-5.5 在 Codex 里测试时，OpenAI 员工已经能明显感到这种风格偏移，于是团队开始追查：这些“哥布林”到底从哪里来？
Read More →
#人工智能
#职业规划
#个人成长
AI 变革与个人应对策略
本文深入探讨了人工智能技术对生产力与社会结构的深刻变革，重点分析了AI时代下的个人生存挑战。文章通过剖析AI驱动的自动化趋势，提出了提升核心竞争力、建立终身学习体系及适应人机协作模式的具体策略，帮助读者在数字化浪潮中保持主动，实现职业发展的转型与升级。
Read More →
#AI Agent
#技术专题
#网络工具
AI-Agent-NetTools: AI 智能体网络交互工具深度解析
本文深入探讨了 AI Agent 在网络交互中的工具应用，详细分析了智能体如何通过特定接口与互联网资源进行交互。内容涵盖了 Agent 设计模式、常用网络工具集及其在自动化工作流中的关键作用，旨在为开发者提供构建高效、智能化网络协作 Agent 的技术参考与实践指导，提升大模型在复杂网络环境下的自主解决问题能力。
Read More →
#Anthropic
#OpenAI
#AI Agent
Anthropic 卖托管运行时，OpenAI 卖可组装底座：Agent 基础设施之争——战略篇
本文深度解析了Anthropic与OpenAI在Agent基础设施领域的战略差异。Anthropic侧重于提供托管运行时环境，强调稳定与易用；OpenAI则致力于构建可组装的底层模型架构，赋予开发者更高灵活度。通过对比两者的技术路线与商业逻辑，本文探讨了AI Agent时代的演进路径及市场竞争焦点，为理解大模型生态的未来布局提供了深刻洞察。
Read More →
#Anthropic
#OpenAI
#Agent
Anthropic 卖托管运行时，OpenAI 卖可组装底座：Agent 基础设施之争——技术篇
本文深入探讨了Agent基础设施的演进趋势，对比了Anthropic通过托管运行时（Computer Use）提供一站式操作能力，与OpenAI通过可组装底座（Swarm等框架）提供灵活开发接口的不同路径。文章通过技术剖析，揭示了双方在AI代理构建模式上的战略差异及对开发者生态的影响，旨在帮助技术人员理解Agent开发范式的变革。
Read More →
No matching articles found
Try adjusting your filter criteria