https://hugozhu.site/tags/agent-spec/Recent content in Agent-Spec on All about Raspberry PiHugoenWed, 19 Aug 2026 00:00:00 +0000https://hugozhu.site/post/2026/357-agent-spec-labor-contract/Wed, 19 Aug 2026 00:00:00 +0000https://hugozhu.site/post/2026/357-agent-spec-labor-contract/<p>前几天有人问我：Agent Builder 的终局是什么？</p> <p>我没直接回答，先让他做了个实验：在 Builder 里输入一句话——「帮我创建一个采购数字员工」。</p> <p>Builder 吐出来一份 YAML：</p> <div class="highlight"><pre tabindex="0" class="chroma"><code class="language-yaml" data-lang="yaml"><span class="line"><span class="cl"><span class="nt">employee</span><span class="p">:</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span><span class="nt">role</span><span class="p">:</span><span class="w"> </span><span class="l">采购专员</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="nt">mission</span><span class="p">:</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span>- <span class="l">处理采购申请</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="nt">context</span><span class="p">:</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span>- <span class="l">ERP</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span>- <span class="l">DingTalk</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span>- <span class="l">Supplier DB</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="nt">permissions</span><span class="p">:</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span>- <span class="l">read_inventory</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span>- <span class="l">create_purchase_draft</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="nt">boundaries</span><span class="p">:</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span>- <span class="l">cannot_approve_purchase</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="nt">evals</span><span class="p">:</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="w"> </span>- <span class="l">...</span><span class="w"> </span></span></span><span class="line"><span class="cl"><span class="c"># generated by hugo AI</span><span class="w"> </span></span></span></code></pre></div><p>我问他：你觉得这份 YAML 是什么？</p> <p>他说：Agent 的配置文件。</p> <p>我说：逐字段再看一遍。</p> <p>他看了几秒，声音不太确定了：「这……是一份 JD？」</p> <p>对。这就是这篇文章的核心判断：<strong>Agent Spec 不是软件描述，它是劳动关系的一份抽象。</strong></p> <p><a href="https://hugozhu.site/img/2026/agent-spec-labor-contract.png"><img src="https://hugozhu.site/img/2026/agent-spec-labor-contract-thumb.jpg" alt="Agent Spec Is the Labor Contract of a Digital Employee"></a></p>https://hugozhu.site/post/2026/358-palantir-ontology-to-agent-we-grow-back/Wed, 19 Aug 2026 00:00:00 +0000https://hugozhu.site/post/2026/358-palantir-ontology-to-agent-we-grow-back/<p>前几天有人抛了一个问题：「哪些客户的续约风险正在上升，今天由谁介入？」</p> <p>我把它原样丢给了一个 AI 助手，想看看它怎么答。</p> <p>它的回答很诚实：答不了。它列出了这个问题需要同时拉齐的五个系统——CRM 里的客户关系、合同系统里的到期日、产品日志里的用量变化、工单里的故障、财务系统里的回款——然后说：「任何一个单独看都不够，五个信号交叉才能定位谁在恶化。」</p> <p>这个回答本身没问题。但它暴露了企业 AI 的真实现状：<strong>AI 不缺智能，缺的是跨系统取数的资格。</strong> 它能看到每一个系统，如果它有资格的话。</p> <p>有意思的是，最近两篇文章，从两个完全不同的方向，撞上了同一个问题。一篇是 Carlos Perez 的<a href="https://medium.com/intuitionmachine/openclaw-is-a-consumer-story-the-enterprise-version-will-be-a-trillion-dollar-reckoning-adafd9a0a87b">企业版 OpenClaw 万亿清算论</a>，一篇是《读懂 Palantir Ontology》的完结篇——后者举的范例决策，一字不差就是上面那个续约风险问题。</p> <p>两篇文章各讲了企业 AI 的一半。把它们拼起来，你会看到一条没人明说的路线。</p> <p><a href="https://hugozhu.site/img/2026/palantir-ontology-to-agent-we-grow-back.png"><img src="https://hugozhu.site/img/2026/palantir-ontology-to-agent-we-grow-back-thumb.jpg" alt="Palantir Grows Agents from Ontology — We Should Grow Ontology Back from Agents"></a></p>