八月一日

Vercel CEO Guillermo Rauch describes the emerging software factory as a continuous Issue → Agent → PR → Release loop, with authors and maintainers increasingly responsible for improving the loop and deciding what deserves attention. Linear Head of Product Nan Yu adds operational evidence: roughly 30% of Linear's bugs already travel through a version of that full flow. Her guidance is deliberately cautious. Agents should investigate root causes extensively, gather evidence through Datadog and Sentry MCPs, ask reporters for missing reproduction steps, and attempt a fix only when confidence is high. The important shift is not merely automating code generation, but encoding sound engineering judgment into the workflow around the model.

Vercel CEO Guillermo Rauch 将正在形成的 software factory 描述为一个持续运转的 Issue → Agent → PR → Release 闭环，作者和维护者的职责则逐渐转向优化这个闭环，并决定哪些问题值得处理。Linear Head of Product Nan Yu 提供了实际运营数据：Linear 大约 30% 的 bug 已经能够走完类似的完整流程。她给出的建议相当克制：agent 应深入调查 root cause，通过 Datadog 和 Sentry MCP 收集证据，在缺少复现步骤时向报告者追问，并且只在置信度足够高时尝试修复。真正重要的变化不只是自动生成代码，而是把可靠的工程判断写进模型周围的工作流。

Box CEO Aaron Levie argues that as agent tasks expand from millions to tens or hundreds of millions of tokens, the harness becomes nearly as important as raw model capability. How work is decomposed and routed can determine both accuracy and cost. Vercel's AI Gateway turns that thesis into infrastructure with budgets by key, team, or project, provider failover, model choice, and real-time observability. YC CEO Garry Tan points to the same layer from another direction, recommending the free, open-source harness his team uses for a personal AI or company brain. Together, the posts suggest that competitive advantage is moving upward from model access toward orchestration, controls, and repeatable operating context.

Box CEO Aaron Levie 认为，当 agent 任务从数百万 token 扩展到数千万甚至数亿 token 时，harness 的重要性将接近模型本身的能力。如何拆解任务、在正确时机路由到正确模型，会直接决定准确率与成本。Vercel 的 AI Gateway 正在把这一判断落实为基础设施，提供按 key、team 或 project 设置预算、provider failover、模型选择和实时 observability。YC CEO Garry Tan 则从另一个方向指向同一层，推荐其团队每天使用的免费开源 harness，可用于 personal AI 或 company brain。这些观点共同说明，竞争优势正在从模型访问权向 orchestration、控制能力与可重复的运行上下文上移。

Zara Zhang says 65% of PRs from Anthropic's product and engineering teams are now raised by Claude Tag, and argues that the natural interface for non-engineering agents is the collaboration software people already use. Her own progression went from terminal in January, to a desktop app in March, to a work collaboration tool in June. Sam Altman offers a consumer version of the same pattern: connect family calendars and children's interests, then have ChatGPT Work create a personalized podcast for the school drive covering a soccer game, an upcoming birthday, and news. Both examples frame agent adoption as an interface and context problem, not simply a model-quality problem.

Zara Zhang 表示，Anthropic 产品与工程团队目前有 65% 的 PR 由 Claude Tag 发起，并认为面向非工程团队的 agent，最自然的入口就是人们已经在使用的协作软件。她自己的使用路径也从 1 月的 terminal，转向 3 月的 desktop app，再到 6 月的工作协作工具。Sam Altman 给出了同一趋势的消费级版本：连接家庭日历并说明孩子们的兴趣，让 ChatGPT Work 每天为上学路程生成一档个性化 podcast，内容包括当天的足球比赛、即将到来的生日和新闻。两个例子都把 agent adoption 定义为界面与上下文问题，而不只是模型质量问题。

OpenAI's Thibault Sottiaux reset usage limits for Codex and ChatGPT Work for the weekend, saying the move followed a week focused on efficiency and inviting users to run 100,000 Luna threads. The temporary reset is another sign that usage capacity and the economics of long-running agent work are becoming visible product levers, alongside model intelligence itself.

OpenAI 的 Thibault Sottiaux 在周末重置了 Codex 与 ChatGPT Work 的使用限额。他表示，此举延续了团队一周以来对效率的投入，并邀请用户运行 10 万个 Luna thread。这次临时重置再次说明，使用容量和长时间 agent 任务的经济性，正与模型 intelligence 一样成为用户可感知的产品杠杆。

Swyx argues that `/loop` and `/goal` remain useful in the GPT-5.6 and Claude 5 era when a task needs both steerability and autonomy, especially for open-ended outcomes where the route should not be exhaustively specified. Peter Steinberger reports a related improvement in model interaction: with version 5.5, he can add requests while the model is working and have them processed without the confusion that previously made a queue necessary. At the specialization end of the spectrum, Replit CEO Amjad Masad highlights an 8B chess model rated around 1500 Elo that consistently beats frontier models and Stockfish level 0 while spending one to two seconds per move rather than 30. These examples point to progress on two axes: better control of general agents and compact models that outperform larger systems inside a narrow task.

Swyx 认为，在 GPT-5.6 与 Claude 5 时代，`/loop` 和 `/goal` 依然有价值，尤其适合同时需要 steerability 与 autonomy、但不应把实现路径规定得过细的开放式任务。Peter Steinberger 则观察到一种相关的交互改进：使用 5.5 时，他可以在模型工作过程中继续追加需求，模型仍能依次处理，不再像以前那样必须依赖 queue 来避免混乱。在 specialization 的另一端，Replit CEO Amjad Masad 展示了一个约 1500 Elo 的 8B 国际象棋模型，它能稳定击败 frontier model 和 Stockfish level 0，而且每步只需 1 至 2 秒，而不是 30 秒。这些案例指向两条同时推进的路线：让通用 agent 更容易控制，以及让小模型在窄任务中超过更大的系统。

Claude Code now supports artifacts in beta for Claude Team and Enterprise organizations through the CLI and desktop app. A session can turn its codebase, connectors, and conversation into a live visual page such as a PR walkthrough, incident timeline, dashboard, release checklist, license audit, or architecture map. Updates publish as new versions at the same link and refresh for viewers, with version history, organization-only sharing, admin controls, retention policies, and compliance API visibility. The feature makes the agent's evolving work product directly inspectable by teammates instead of requiring a separate status reconstruction.

Claude Code 现已通过 CLI 和 desktop app，面向 Claude Team 与 Enterprise 组织提供 artifacts beta。一个 session 可以把 codebase、connector 与对话上下文转化为实时可视页面，例如 PR walkthrough、incident timeline、dashboard、release checklist、license audit 或 architecture map。更新会在同一链接发布为新版本并自动刷新给查看者，同时保留 version history、组织内分享、管理员控制、retention policy 和 compliance API 可见性。这个功能让团队成员能够直接检查 agent 持续演进的工作成果，而不必另行重建一份状态汇报。

The Takeaway: Igor Babushkin says coding agents became impossible for software engineers to ignore around November and December, but coding is only the first tractable domain because its outcomes are verifiable. He expects rapid progress in coding, math, and scientific work where tools such as Lean or real-world experiments can close the reward loop. Moving into long-horizon, non-verifiable domains is harder: a 24-hour agent rollout makes every reinforcement-learning update slow, although he believes increasingly capable models may enable useful LLM judges over the coming months or year.

His new company River AI is pursuing three bets. River API provides reinforcement learning and fine-tuning intended to be cheap, reliable, and scalable; its personalization research challenges the practice of optimizing one model for the average user; and its hardware work asks whether frontier inference can run locally for lower latency, richer voice and video interaction, privacy, and user control. Babushkin frames distribution of AI's benefits and control as an urgent safety issue, and predicts a bifurcation between expensive super-AI available to a few and everyday personal AI optimized to improve an individual's life.

Babushkin also sees pressure on proprietary model vendors. Frontier pretraining has diminishing returns while open models keep improving, so he argues closed providers must find fundamental innovations, new domains, and new revenue paths. Companies with unique data and domain expertise may eventually prefer post-training their own models rather than handing away the foundation of their advantage. In his view, the harness, proprietary data, and distributed domain expertise become more important as simply scaling a general model gets harder.

核心结论： Igor Babushkin 表示，大约从 11 月和 12 月开始，coding agent 已经强到 software engineer 无法忽视，但 coding 只是第一个容易突破的领域，因为结果可以验证。他预计 coding、数学，以及能够通过 Lean 或现实实验闭合 reward loop 的科学任务会快速进步。进入 long-horizon、non-verifiable 领域则困难得多：如果一次 agent rollout 需要 24 小时，每次 reinforcement learning 更新都会非常缓慢。不过，他认为随着模型判断能力提升，未来几个月到一年内，LLM judge 可能开始产生实际效果。

他的新公司 River AI 正在押注三个方向。River API 提供 reinforcement learning 与 fine-tuning 服务，目标是做到低成本、可靠且可扩展；personalization 研究挑战了为“平均用户”优化同一个模型的惯例；hardware 项目则探索能否在本地运行 frontier inference，从而获得更低 latency、更丰富的语音和视频交互、更强隐私与用户控制。Babushkin 将 AI 收益与控制权的分散视为一项紧迫的 safety 议题，并预测市场会分化为两类：少数人可以使用的昂贵 super-AI，以及以改善个人生活为目标的日常 personal AI。

Babushkin 同时认为 proprietary model provider 正面临压力。Frontier pretraining 的回报递减，而 open model 持续追赶，因此 closed provider 必须找到基础性创新、新领域和新的 revenue 路径。拥有独特数据与 domain expertise 的公司，最终可能更愿意自行 post-train 模型，而不是把自身竞争优势的基础交给外部供应商。在他看来，当单纯扩大通用模型越来越难时，harness、专有数据和分布式领域知识会变得更加重要。