AI Builders Digest — 2026-06-09

X / TWITTER

Boris Cherny, Claude Code at Anthropic

Boris Cherny says long-running autonomous work is where Opus is currently standing out. His operating recipe is practical: use auto mode for permissions, give Claude dynamic workflows so it can orchestrate many agents, use /goal or /loop to keep momentum, run Claude Code in the cloud, and make sure the agent can verify its own work end to end through browser, simulator, or service-level checks.

Boris Cherny 认为，长时间自主执行任务是 Opus 目前表现突出的场景。他给出的操作建议很具体：权限使用 auto mode，让 Claude 能通过 dynamic workflows 编排大量 agents，用 /goal 或 /loop 保持任务推进，在 cloud 里运行 Claude Code，并且给 agent 配好端到端自检能力，比如 browser、mobile simulator 或后端服务启动验证。

Source: https://x.com/bcherny/status/2063792263067754658

Thibault Sottiaux, Codex and ChatGPT at OpenAI

Thibault Sottiaux announced a 100-day Codex usage experiment: each day, OpenAI will pick one person doing impressive or useful work with Codex and give them 10x usage limits for a month. The signal is clear: usage ceilings are becoming an active product lever, not just a pricing constraint.

Thibault Sottiaux 宣布了一个为期 100 天的 Codex 使用实验：每天挑选一位用 Codex 做出出色或实用成果的人，给予一个月 10 倍使用额度。这里的信号很明确：usage limit 正在从单纯的价格约束，变成产品增长和生态激励手段。

Source: https://x.com/thsottiaux/status/2063748242681307611

Madhu Guru, former Google Gemini / Veo product leader

Madhu Guru pushes back on the idea that training data is low-skill labeling work. Frontier progress increasingly depends on high-economic-value task data, especially outside software engineering, where knowledge is domain-specific, poorly documented, and buried across legacy tools. His point: the reason SWE agents arrived first is not only model capability, but the unusually legible structure of software work.

Madhu Guru 反驳了“训练数据只是低技能标注”的看法。模型前沿进展越来越依赖高经济价值任务数据，尤其是在非软件工程领域，这些知识高度垂直、缺少文档，还分散在各种 legacy tools 里。他的核心判断是：SWE agents 先跑出来，不只是因为模型更擅长写代码，也因为软件工作的结构化程度更高。

Source: https://x.com/realmadhuguru/status/2063704354910347520

Guillermo Rauch, Vercel CEO

Guillermo Rauch says Vercel AI Gateway recovers more than 1 trillion tokens per month on average, comparing it to Stripe-style recovery for failed payments. The value proposition is not cheaper tokens alone, but routing, redundancy, zero-data-retention enforcement, observability, usage APIs, and caps without markup over model labs.

Guillermo Rauch 表示，Vercel AI Gateway 平均每月恢复超过 1 万亿 tokens，类似 Stripe 对失败支付的智能重试。它卖的不是单纯“更便宜的 token”，而是 routing、冗余、zero-data-retention enforcement、可观测性、usage APIs 和 caps 这些 AI 基础设施能力，并且不在模型厂商价格上加价。

Source: https://x.com/rauchg/status/2063714700618334260

Aaron Levie, Box CEO

Aaron Levie argues that AI workloads will split across model families: frontier models for high-end tasks and cheaper models for high-volume work. That makes routing increasingly valuable, because agent orchestration must optimize cost while still completing the task successfully.

Aaron Levie 认为，未来一两年 AI use cases 会在不同模型家族之间分层：高端任务使用 frontier models，大量高频任务转向更便宜的模型。这会让 routing layer 变得更有价值，因为 agent orchestration 的难点会变成：在保证任务完成质量的同时优化成本。

Source: https://x.com/levie/status/2063835799096090749

Aaron Levie also argues that AI has not erased the hard parts of enterprise software. Even if development gets cheaper, enterprise GTM, consultative selling, implementation, security, integration, and differentiation become more important because buyers face an even noisier market.

Aaron Levie 还指出，AI 并没有消除 enterprise software 真正难的部分。即使软件开发变便宜，企业级 GTM、顾问式销售、实施、安全、集成和差异化反而更重要，因为买方面对的市场会更拥挤、更难判断。

Source: https://x.com/levie/status/2063756386572681606

Box also shipped a web markdown editor with CLI support, commenting, version history, and mounted-drive access through Box Drive for tools like Claude, Codex, Obsidian, and Cursor.

Box 同时上线了 web markdown editor，支持 CLI、评论、版本历史，并可通过 Box Drive 挂载到本地，让 Claude、Codex、Obsidian、Cursor 等工具直接处理 Box 文件。

Source: https://x.com/levie/status/2063649508681224367

Garry Tan, Y Combinator President and CEO

Garry Tan says teaching people how to use AI tools has become a serious bottleneck, pointing to YC's own work on this problem. He also shared a GBrain update that summarizes how a founder's thinking has changed over time, suggesting a direction where AI tools become memory and reflection infrastructure, not just task executors.

Garry Tan 认为，教会人们真正使用 AI tools 已经成为严重瓶颈，并指向 YC 在这方面的投入。他还分享了 GBrain 更新：可以总结一个人的思考随时间如何变化。这说明 AI 工具正在从“任务执行器”延伸到记忆、复盘和认知演化基础设施。

Sources: https://x.com/garrytan/status/2063786111588323780, https://x.com/garrytan/status/2063786182140735829, https://x.com/garrytan/status/2063785286367392095

Zara Zhang, builder

Zara Zhang says her Frontend Slides skill has grown organically because slides are inherently social: people see the output, ask how it was made, and treat HTML decks as a visible marker of being AI-native. The useful product lesson is that some AI-native workflows spread because their artifacts advertise the workflow.

Zara Zhang 认为，她的 Frontend Slides skill 能自然增长，是因为 slides 天生具备社交传播属性：别人看到成品，会问“这是怎么做的”，并把 HTML decks 视为 AI-native 能力的外显标志。这里的产品启发是：有些 AI-native workflow 会因为产物本身能展示方法而自传播。

Source: https://x.com/zarazhangrui/status/2063638307586662539

Nikunj Kothari, FPV Ventures partner

Nikunj Kothari argues that companies should give employees generous token budgets so they can stay near the frontier and explore edge cases. His warning: token optimization can become premature efficiency if it pushes people back into old workflows too early.

Nikunj Kothari 认为，公司应该给员工足够多的 token budget，让他们持续接近前沿并探索边界场景。他的提醒是：token optimization 如果过早发生，可能会变成“效率之名下的保守”，让人重新退回旧工作流。

Source: https://x.com/nikunj/status/2063630238123483195

Peter Steinberger, OpenClaw and OpenAI

Peter Steinberger's concise framing landed widely: users should stop thinking of coding agents as things to prompt directly, and start designing loops that prompt agents. This is the same shift showing up across builder conversations: from single-shot prompt craft to durable agent operating systems.

Peter Steinberger 的一句话获得很高传播：不要再把 coding agents 当成一次性 prompt 的对象，而要设计会 prompt agents 的 loops。这与今天多个 builder 的共识一致：重点正在从单次 prompt 技巧转向可持续运行的 agent operating systems。

Source: https://x.com/steipete/status/2063697162748260627

Aditya Agarwal, South Park Commons general partner and Bevel Health co-founder

Aditya Agarwal reflects on post-IPO liquidity from his Meta and Dropbox experience: wealth tends to amplify deeper desires rather than create new ones. For Silicon Valley, upcoming liquidity events may recycle talent and capital into new companies, funds, and ambitious experiments.

Aditya Agarwal 结合自己经历 Meta 和 Dropbox IPO 的经验说，财富通常不是创造新欲望，而是放大更深层的欲望。对 Silicon Valley 来说，接下来一波流动性事件可能会把人才和资本重新投入新公司、新基金和更激进的实验。

Source: https://x.com/adityaag/status/2063731771284619521

PODCASTS

The MAD Podcast with Matt Turck — State of Enterprise AI 2026: Aaron Levie on Tokenmaxxing, Rise of Headless, and AI-Proofing Your Job

The Takeaway: Aaron Levie sees enterprise AI adoption as real and optimistic, but slowed by a paradox: model capability keeps improving so fast that rollout plans become obsolete before companies finish implementing them.

Aaron Levie, CEO of Box, is worth listening to because he sits between Silicon Valley's agent-native builders and large enterprise buyers who still have security, integration, and change-management constraints. His sharpest framing is that the gap is now less "Silicon Valley vs. everyone else" and more "AI-native engineering teams vs. non-engineering knowledge work." Enterprises have mostly learned how to roll out chat-based AI, but the next wave is agentic work that actually performs tasks inside workflows. That requires cost controls, permission models, evaluation, internal implementation talent, and business-process redesign.

Levie's most practical point is that AI progress can slow enterprise adoption because each new capability changes the architecture buyers thought they had just standardized. He also ties token anxiety to workload routing: expensive frontier models will be justified for high-value work, while cheaper models will handle high-volume tasks. The companies that can orchestrate this split securely and cheaply will sit in a valuable layer of the stack.

核心判断：Aaron Levie 认为 enterprise AI adoption 真实且乐观，但会被一个悖论拖慢：模型能力进步太快，企业还没完成部署，原来的 rollout plan 就已经过时。

Box CEO Aaron Levie 的视角有价值，因为他正好站在两端之间：一端是 Silicon Valley 的 agent-native builders，另一端是仍然受安全、集成、变更管理约束的大型企业买家。他最锋利的判断是，现在的差距不再只是“Silicon Valley vs. 其他人”，而是“AI-native 工程团队 vs. 非工程知识工作”。企业基本刚学会部署 chat-based AI，下一波却已经变成能在 workflow 里真正做事的 agentic work。这需要成本控制、权限模型、评估体系、内部实施人才和业务流程重构。

Levie 最实用的提醒是：AI 进步本身可能拖慢企业采用，因为每一次新能力都会改变买方刚刚准备标准化的架构。他也把 token anxiety 和 workload routing 联系起来：高价值任务会继续使用昂贵的 frontier models，高频任务会被更便宜的模型承接。能安全、低成本地编排这种分层的公司，会占据 AI stack 中很有价值的一层。

Source: https://www.youtube.com/watch?v=Gs2styCcwro

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