55个点赞，19条评论，登Hacker News首页前20。
这是Herdr上线第一天的数据。一个跑在终端里的Agent多路复用器——简单说，你可以在命令行里同时调用多个AI Agent（Claude Code、Cursor Agent、Codex等），让它们并行处理同一个任务的不同部分，然后对比结果择优采纳。像指挥一个AI团队一样指挥你的终端。
Herdr不是这周唯一冒出来的Agent工具。
AWS发布了开源框架Blocks，定位是"面向AI智能体的后端开发工具"——帮开发者快速构建Agent的状态管理、工具调用、记忆检索等后端基础设施。量子位报道了BrowserBC：人类在网页上操作一次，系统就能克隆出对应的Agent行为模式，一次录制所有Agent学会。GitLab 19.0直接把Agentic AI嵌入了凭证管理、合并请求和供应链安全——AI不再是代码补全工具，而是变成了开发流程的原生参与者。Lore项目在HN上获得29个点赞，它的功能是把团队的技术决策喂给coding agent，让AI理解"为什么这么做"而不仅仅是"代码是什么"。
这些产品集中在一周内涌现，不是巧合。
从Prompt工程到工具链工程
2024-2025年，Agent开发的主流范式是"Prompt工程"——写一段好的system prompt，加几轮few-shot示例，再配上function calling，一个Agent就"能用"了。但"能用"和"好用"之间隔着一条巨大的鸿沟：Agent经常在多步骤任务中跑偏，不会处理异常状态，无法复用之前的经验，团队协作场景下更是一团糟。
2026年中，Agent开发正在进入一个新的阶段：工程化。开发者们开始意识到，Agent不是一个更聪明的Chatbot，而是一个需要完整工具链支撑的新型软件形态。它需要版本控制（Lore）、需要多路调度（Herdr）、需要行为克隆（BrowserBC）、需要后端框架（Blocks）、需要深度集成进开发流程（GitLab 19.0）。
这个转变和移动互联网早期的发展轨迹很像。2008年App Store刚上线时，一个人花一周就能写出一个iOS App；但到2015年，一个正经的移动应用需要完整的CI/CD、A/B测试、埋点分析、崩溃监控、推送系统——基础设施越来越厚，门槛越来越高，但产品质量也越来越可靠。
Agent的"操作系统"正在形成
如果把这些工具拼在一起看，你会发现一个Agent开发的"操作系统"轮廓正在浮现：
底层是后端框架（AWS Blocks），处理状态、记忆、工具调用等基础能力；中间是编排和调度层（Herdr），负责多个Agent的分工协作和结果比对；上层是能力获取层（BrowserBC），让Agent通过观察人类快速习得新技能；集成层（GitLab 19.0）把Agent嵌入已有工作流；知识层（Lore）让Agent理解组织上下文。
MCP（Model Context Protocol）在这个体系中扮演了"USB-C接口"的角色——Anthropic推出的MCP协议月下载量已达9700万次，成为Agent工具互联的事实标准。当Agent有了标准的接口、完整的工具链、可靠的工程实践，"生产级Agent"才从一个口号变成现实。
"2024年是Agent的'Hello World'年，2025年是'Demo年'，2026年是'生产年'。区别在于：生产年的标志不是更多Demo，而是更多基础设施。"
—— API7.ai创始人温铭在InfoQ访谈中的判断
独立开发者的机会窗口
值得注意的是，这批Agent工具链产品大多来自独立开发者和小团队，而非大公司。Herdr是个人项目，BrowserBC来自小团队，Lore也是开源社区项目。
这符合一个规律：基础设施的范式转移期，往往是小团队创新最活跃的窗口。大公司有存量包袱——AWS当然可以做Blocks，但它不会做一个帮你同时调度多个竞争对手模型的工具（Herdr）；GitLab可以把Agent嵌入合并请求，但它不会做一个跨平台的Agent行为克隆工具。
API7.ai创始人温铭最近总结了烧了几百亿Token用AI重写生产级网关的6条经验，其中一条是："Agent不是要替代工程师，而是要求工程师升级为Agent的管理者和教练。"当工具链越来越完善，一个人指挥一个Agent团队完成复杂任务的时代正在到来。
但别忘了，工具只是工具。17倍错误放大的研究数据还在耳边——86%的团队用了Agent，40%在半年内失败。工具链降低了入门门槛，但能不能用好Agent，归根结底取决于使用工具的人是否理解问题本身。
55 upvotes, 19 comments, on the Hacker News front page top 20.
That's Herdr's first day data. An Agent multiplexer that lives in your terminal -- simply put, you can call multiple AI Agents (Claude Code, Cursor Agent, Codex, etc.) simultaneously from the command line, have them work in parallel on different parts of the same task, then compare results and pick the best. Command your terminal like you're leading an AI team.
Herdr wasn't the only Agent tool popping up this week.
AWS released Blocks, an open-source framework positioned as a "backend development tool for AI Agents" -- helping developers quickly build Agent backend infrastructure for state management, tool calling, memory retrieval, etc. QbitAI reported on BrowserBC: a human operates a web page once, and the system clones the corresponding Agent behavior pattern; one recording, all Agents learn. GitLab 19.0 embedded Agentic AI directly into credential management, merge requests, and supply chain security -- AI is no longer a code completion tool but a native participant in the development process. The Lore project got 29 upvotes on HN; its function is feeding a team's technical decisions to coding agents so AI understands "why it was done this way" not just "what the code is."
These products all emerging within one week isn't a coincidence.
From Prompt Engineering to Toolchain Engineering
In 2024-2025, the dominant Agent development paradigm was "Prompt engineering" -- write a good system prompt, add a few few-shot examples, pair with function calling, and an Agent is "usable." But between "usable" and "good enough" lies a massive chasm: Agents frequently go off track in multi-step tasks, don't handle exceptions, can't reuse previous experience, and are a mess in team collaboration scenarios.
By mid-2026, Agent development is entering a new phase: engineering. Developers are beginning to realize that Agents aren't smarter Chatbots -- they're a new form of software requiring a complete toolchain. They need version control (Lore), multi-path scheduling (Herdr), behavior cloning (BrowserBC), backend frameworks (Blocks), and deep integration into development workflows (GitLab 19.0).
This shift resembles the early mobile internet trajectory. When the App Store launched in 2008, one person could write an iOS app in a week; but by 2015, a serious mobile app needed full CI/CD, A/B testing, analytics, crash monitoring, push systems -- infrastructure thickens, barriers rise, but product quality becomes far more reliable.
An Agent "Operating System" Is Taking Shape
If you piece these tools together, the outline of an Agent development "operating system" is emerging:
At the bottom is the backend framework (AWS Blocks), handling basics like state, memory, and tool calls; in the middle is orchestration and scheduling (Herdr), responsible for multi-Agent division of labor and result comparison; above that is capability acquisition (BrowserBC), letting Agents quickly learn new skills by observing humans; the integration layer (GitLab 19.0) embeds Agents into existing workflows; the knowledge layer (Lore) lets Agents understand organizational context.
MCP (Model Context Protocol) plays the role of "USB-C port" in this system -- Anthropic's MCP protocol has reached 97 million monthly downloads, becoming the de facto standard for Agent tool interconnection. When Agents have standard interfaces, complete toolchains, and reliable engineering practices, "production-grade Agents" goes from slogan to reality.
"2024 was Agent's 'Hello World' year; 2025 was the 'Demo year'; 2026 is the 'production year.' The difference: the mark of a production year isn't more demos -- it's more infrastructure."
-- Wen Ming, API7.ai founder, in an InfoQ interview
Indie Developers' Window of Opportunity
Notably, this wave of Agent toolchain products mostly comes from independent developers and small teams, not big companies. Herdr is a personal project; BrowserBC comes from a small team; Lore is an open-source community project.
This follows a pattern: during infrastructure paradigm shifts, small-team innovation tends to be most active. Big companies have legacy baggage -- AWS can build Blocks, but it won't build a tool that helps you orchestrate multiple competitors' models simultaneously (Herdr); GitLab can embed Agents in merge requests, but it won't build a cross-platform Agent behavior cloning tool.
API7.ai founder Wen Ming recently summarized 6 lessons from burning tens of billions of Tokens rewriting production-grade gateways with AI, one being: "Agents aren't here to replace engineers; they require engineers to level up into Agent managers and coaches." As toolchains mature, the era of one person commanding an Agent team to complete complex tasks is arriving.
But don't forget: tools are just tools. The research finding of 17x error amplification still rings in our ears -- 86% of teams that adopted Agents saw 40% fail within six months. Toolchains lower the barrier to entry, but whether you can use Agents well ultimately depends on whether the person using the tool understands the problem itself.
Agent Toolchain · AWS Blocks · Herdr · BrowserBC · Lore · Agent Engineering · MCP Protocol · From Prompts to Engineering
