2026-04-14

AI Builders Digest — 2026-04-14

X / TWITTER

Guillermo Rauch (CEO @Vercel)

Vercel open sourced Open Agents (github.com/vercel/open-agents), a reference platform for building cloud coding agents. The thesis: the competitive moat of software companies is shifting from "the code they wrote" to "the means of production of that code" — i.e., your AI factory. Built on Vercel's agentic infrastructure: Fluid for running the agent brain, Workflow for long-running durability, Sandbox for secure code execution, and AI Gateway for multi-model token routing. Companies like Stripe, Ramp, Spotify, and Block are already building their own versions.

"Off-the-shelf coding agents don't perform well with huge monorepos, don't have your institutional knowledge, integrations, and custom workflows."

来源: https://x.com/rauchg/status/2043869656931529034

Vercel CEO Guillermo Rauch 开源了 Open Agents (github.com/vercel/open-agents)，一个构建云端 coding agent 的参考平台。他的核心观点是：软件公司的护城河正在从"他们写的代码"转向"代码的生产方式"——也就是你的 AI 工厂。该项目基于 Vercel 的 agent 基础设施：Fluid 运行 agent 大脑、Workflow 处理长时间运行任务、Sandbox 确保代码执行安全、AI Gateway 管理多模型 token 路由。Stripe、Ramp、Spotify、Block 等公司已经在自建类似系统。

Aaron Levie (CEO @Box)

Levie predicts a new enterprise role is emerging: the agent deployer and manager. Their job: identify high-leverage workflows where agents can 100x throughput, connect business systems, design human-agent handoffs, manage evals, and track KPIs. This person needs to be technical enough for skills, MCP, and CLIs — but also operationally strong. Levie sees this as either an existing employee repositioned or a net-new hire, likely one per team, reporting into AI or IT. He also notes it's an ideal role for next-gen technical hires leaning into AI.

"The person must be good at mapping the process and understanding where the value could be unlocked and be relatively technical… But also great operationally and at business."

来源: https://x.com/levie/status/2043883641366032638

Box CEO Aaron Levie 预判企业里将出现一个新岗位：Agent 部署与管理专家。这个人的职责是：找出哪些工作流可以让 agent 带来 100 倍效率提升，连接业务系统，设计人机协作节点，管理评估体系，追踪 KPI。既要懂技术（skills、MCP、CLI），又要有运营和商业思维。Levie 认为这个角色可以由现有员工转型，也可以全新招聘，最终可能是每个团队配一个，向 AI 或 IT 部门汇报。对下一代技术背景的 AI 爱好者来说，这是个理想的职业方向。

Thariq (Claude Code @Anthropic)

Thariq shared that the new Claude Code renderer is live. Enable it with CLAUDE_CODE_NO_FLICKER=1 claude. The tweet got 1,043 likes — a strong signal of how hungry the developer community is for better AI coding UX. Docs available at the linked URL.

来源: https://x.com/trq212/status/2043814646600348046

Anthropic Claude Code 团队发布了全新版 Claude Code 渲染器，可通过 CLAUDE_CODE_NO_FLICKER=1 claude 开启。该推文获得 1,043 个点赞，说明开发者社区对更好的 AI 编程体验需求强烈。完整文档见链接。

Peter Yang (Product @Roblox)

Yang raises an uncomfortable point: the GPT integration in OpenClaw may be at risk of falling behind. His take: "The bar is GPT needs to be just as good if not better for OpenClaw as Opus." Translation — if OpenAI's models can't match what Anthropic's Opus delivers inside OpenClaw, users will migrate. Worth watching for OpenAI's next model move.

来源: https://x.com/petergyang/status/2043732939817791979

Roblox 产品经理 Peter Yang 提出了一个尴尬的问题：OpenClaw 中 GPT 集成的质量可能正在落后。他的判断是："GPT 需要在 OpenClaw 中至少和 Opus 一样好，否则就会失去用户。"换句话说，如果 OpenAI 的模型无法匹敌 Anthropic Opus 在 OpenClaw 中的表现，用户就会转向其他选择。这值得关注 OpenAI 下一步的模型动作。

Amjad Masad (CEO @Replit)

Masad fires two shots. First, a product update: Replit now supports configurable app hosting regions, critical for compliance and privacy law requirements. Second, a sharper observation — a thread-worthy one about how people treat measurability inconsistently: "Most people are loose with what's measurable and rigid with what's inherently fuzzy. They bend data to fit the story, but demand precision from things that don't survive contact with reality. LARPing rigor."

来源: https://x.com/amasad/status/2043785145606656223
来源: https://x.com/amasad/status/2043948524908490953

Replit CEO Amjad Masad 发了两条动态。一是产品更新：Replit 现在支持可配置的 App 托管区域，这对合规和隐私法规很关键。二是更犀利的内容：关于人们对待可量化事物的双重标准——"大多数人对可测量的东西很随意，对本质模糊的东西却很死板。他们扭曲数据来迎合叙事，却对那些经不起现实检验的事物要求精确。这叫 LARPing rigor（装模作样的严谨）。"

Dan Shipper (CEO @Every)

Shipper offers a framework for 2026 software engineering: you need two roles — a pirate and an architect. Pirates move fast, write messy code, figure out what's valuable. Architects turn that mess into well-oiled systems. Also highlighted Kieran's "an agent is just a folder" mental model as genuinely changed how he thinks about AI systems — worth reading the full thread.

来源: https://x.com/danshipper/status/2043819933675450455
来源: https://x.com/danshipper/status/2043739805276619223

Every CEO Dan Shipper 提出了 2026 年软件工程的新框架：需要两种角色——海盗和建筑师。海盗快速行动、写粗糙代码、搞清楚什么有价值。建筑师把那堆乱码变成精密系统。另外他还推荐了 Kieran 提出的"agent 就是一个文件夹"这个 mental model，说这真正改变了他对 AI 系统的认知方式，值得看完整 thread。

Nikunj Kothari (Partner @FPV Ventures)

Kothari with a darkly funny observation about talent dynamics at frontier AI labs: "Frontier labs greatest retention perk must be truly unlimited tokens 😅 Person joins a fast growing 'agentic' startup → discovers there are token limits → discovers they have to use open source models to balance costs → decides to say f this and goes back to said frontier lab." True story, he says. Which means the real perk of working at OpenAI/Anthropic isn't the salary — it's the GPU.

来源: https://x.com/nikunj/status/2043808491807396059

FPV Ventures 合伙人 Nikunj Kothari 用一个略带讽刺的观察揭示了前沿 AI 公司的人才动态："前沿实验室最大的留存福利一定是真正无限的 tokens 😅 一个人加入一家快速增长的'agent'创业公司 → 发现有 token 限制 → 发现不得不使用开源模型来平衡成本 → 决定滚回原来的前沿实验室。"他说这是真实故事。这意味着在 OpenAI/Anthropic 工作的真正福利不是工资——是 GPU。

Swyx

Shared that Anh (a devtools heavy hitter on HN and beyond) open sourced her writing Skills template. Also noted that ~80% of the world's agents and AI engineering is happening within roughly 3 square miles — a vivid way of pointing at the geographic concentration of the current AI build-out.

来源: https://x.com/swyx/status/2043786360012845509
来源: https://x.com/swyx/status/2043778767798317349

Swyx 分享了知名 devtools KOL Anh 开源的写作 Skills 模板。他还指出，全球约 80% 的 agent 和 AI 工程都集中在约 3 平方英里内——形象地点出了当前 AI 建设的地理集中度。

Garry Tan (President & CEO @Y Combinator)

GBrain v0.9.3 is out with search tuning, CJK query improvements, search evals, better health checks, and security hotfixes. YC's own OpenClaw/Hermes-based agent setup is maturing quickly. Tan describes it as "an opinionated setup that is literally my OpenClaw/Hermes Agent setup" — including a voice agent built on OpenAI Realtime (soon upgrading to Gemini Live).

来源: https://x.com/garrytan/status/2043948627291451625
来源: https://x.com/garrytan/status/2043948971568312473

Y Combinator 总裁 Garry Tan 发布 GBrain v0.9.3，包含搜索调优、中日韩查询改进、搜索评估、健康检查优化和安全修复。YC 自研的 OpenClaw/Hermes agent 配置正在快速成熟。Tan 形容这是"一个有态度的配置，就是我个人的 OpenClaw/Hermes Agent 配置"——还包括一个基于 OpenAI Realtime 构建的语音 agent（即将升级到 Gemini Live）。

PODCASTS

No Priors Podcast — "AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus"

Liam Fedus left his role as VP of Post-Training at OpenAI — where he helped ship GPT-4 and early ChatGPT — to co-found Periodic Labs, a company with a simple but audacious mission: build an AI foundation lab for atoms. The core insight driving him: language models changed the digital world, but the real next wave is connecting AI to the physical world — materials science, chemistry, process engineering. He believes the AI of 2022 was simply too weak to make this viable. Now, with stronger models, test-time inference, and reliable tool use, the timing is finally right.

The most compelling concept in the conversation: closed-loop scientific discovery. Periodic's system generates hypotheses, runs physical experiments, learns from results, and iterates — with AI orchestrating specialized atomic models as tools. Fedus draws a direct parallel to how software engineering agents work: write code, run tests, fix bugs, repeat. He expects the same recursive self-improvement dynamic to eventually apply to AI research itself — the outer loop just runs slower because you need GPUs and real experimental cycles instead of instant unit tests.

On robotics: not required for Periodic's core goal, but a massive accelerator once you have dexterous, reliable systems that can operate in unstructured lab environments. Fedus expects humanoid robotics to be a key unlock for spinning up new labs quickly.

His closing vision: if the physical world could evolve at even a fraction of the speed of the digital world, "life will just feel dramatically different." He's essentially betting the next decade belongs to atoms, not bits.

"Science ultimately isn't sitting in a room thinking really hard. You have to conduct experiments. You have to learn from them. You have to interface with reality."

来源: https://www.youtube.com/@NoPriorsPodcast

No Priors Podcast — "AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus"

Liam Fedus 离开 OpenAI VP of Post-Training 的岗位（曾参与 GPT-4 和早期 ChatGPT 的发布），共同创立了 Periodic Labs。这家公司的使命简单却大胆：在原子层面构建 AI 基础实验室。他的核心洞察是：语言模型改变了数字世界，但真正的下一波浪潮来自 AI 与物理世界的连接——材料科学、化学、过程工程。他认为 2022 年的 AI 能力太弱，无法实现这个愿景。现在，有了更强的模型、测试时推理和可靠的工具调用，时机终于成熟。

对话中最吸引人的概念：闭环科学发现。Periodic 的系统生成假设、运行物理实验、从结果中学习、迭代优化——AI 作为编排层，协调专门的原子模型作为工具。Fedus 直接类比软件工程 agent 的工作方式：写代码、跑测试、修 bug、重复。他预计同样的递归自我改进逻辑最终会应用到 AI 研究本身——只不过外层循环更慢，因为需要 GPU 和真实的实验周期，而不是即时的单元测试。

关于机器人：不是 Periodic 核心目标的必需品，但一旦有了灵巧、可靠的机器人系统（在非结构化实验室环境中运行），将成为巨大的加速器。Fedus 预计人形机器人将是快速建立新实验室的关键解锁。

他的终极愿景：如果物理世界能以数字世界速度的一小部分演进，"生活将会感觉完全不同。"他基本上在赌：下一个十年属于原子，而非比特。

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders