八月二十一日

Anthropic's Thariq said the company is rolling out new enterprise safeguards for customers that want Claude to run on infrastructure they control, with tighter control over where data lives and who can access it. The notable point is not just the feature list but the deployment posture: Anthropic is framing agent adoption as something that has to fit existing enterprise boundaries, and he said the work was shaped alongside roughly 100 companies before a broader fall rollout.

Anthropic 的 Thariq 表示，公司正在推出一套新的 enterprise safeguards，面向那些希望让 Claude 运行在自有基础设施上的客户，并让他们更严格地控制数据驻留位置和访问权限。这里真正重要的不只是功能本身，而是部署姿态：Anthropic 正把 agent adoption 定义为必须适配现有 enterprise boundary 的事情；他还表示，这项能力已与大约 100 家公司共同打磨，并计划在秋季更广泛推出。

Boris Cherny said Anthropic has been working with customers on additional safety, privacy, and compliance controls for what he called Mythos-class models. His concrete claim is that customers will be able to own and control their own data while Anthropic retains none, which makes this less about consumer convenience and more about whether frontier agents can clear enterprise governance requirements.

Anthropic 的 Boris Cherny 表示，团队一直在与客户合作，为他所称的 Mythos-class models 增加额外的 safety、privacy 与 compliance controls。他给出的核心说法是：客户可以拥有并控制自己的数据，而 Anthropic 不保留任何数据；这意味着重点不在消费级便利性，而在于 frontier agents 能否满足 enterprise governance 的要求。

OpenAI's Thibault Sottiaux said reports of different Codex usage limits often traced back to `sub2api`, which turns a subscription into re-served API traffic and can trigger fraud-prevention systems. He contrasted that with supported Sign in with ChatGPT usage in official clients and open-source clients such as Pi and OpenCode, then separately pointed to two new creation surfaces: transparent-image generation in GPT-Image-2 and collaborative ChatGPT Sites.

OpenAI 的 Thibault Sottiaux 表示，部分关于 Codex usage limits 差异的反馈，最后都追溯到 `sub2api` 这类把 subscription 转成再分发 API traffic 的做法，因此会触发 fraud-prevention systems。他把这种方式与受支持的 Sign in with ChatGPT 用法区分开来，后者可用于官方客户端以及 Pi、OpenCode 等 open-source clients；与此同时，他还单独强调了两个新的 creation surfaces：GPT-Image-2 的 transparent image generation，以及可协作的 ChatGPT Sites。

Madhu Guru's most useful point is that enterprise teams do not need one grand eval so much as a ladder of evals with different costs and realism. He separates hill-climb evals for pushing capability, regression evals for protecting the current product, smoke tests for things that cannot fail, and launch evals that approximate live traffic, which is a practical framing for teams that still treat evaluation as a single scoreboard.

Meta AI 负责人 Madhu Guru 最有价值的观点是，enterprise teams 真正需要的不是一个“总评测”，而是一套成本与真实性不同的 eval ladder。他把评测拆成用于推进能力边界的 hill-climb evals、保护现有产品的 regression evals、覆盖绝不能出错项的 smoke tests，以及逼近真实流量的 launch evals；对于仍把 evaluation 当作单一记分牌的团队来说，这是一个非常实用的框架。

Aaron Levie argued that companies sitting close to high-volume enterprise workflows can justify post-training models around those tasks, both to raise accuracy and to lower inference cost. The important nuance is his economic test: this only makes sense when the workflow is repeated enough and understood deeply enough that reward shaping and task-specific optimization beat simply using a general frontier model.

Box CEO Aaron Levie 认为，那些贴近高频 enterprise workflows 的公司，可以围绕这些任务去做 post-training，从而同时提升准确率并降低 inference cost。这里的关键细节是他的经济判断：只有当某个 workflow 足够重复、公司对它又理解得足够深时，reward shaping 与 task-specific optimization 才会比直接调用通用 frontier model 更划算。

Peter Yang floated a simple but useful hypothesis: a meaningful share of mediocre AI output may be improved by having one agent push another to review its work and try again. It is not a measured result, but it is a good reminder that teams should test a cheap manager-worker critique loop before they assume quality problems require a much heavier orchestration stack.

Peter Yang 提出了一个简单但有启发性的假设：不少平庸的 AI output，也许只需要让一个 agent 去督促另一个 agent 复查并重做，就能明显改善。它还不是一个经过验证的结果，但它提醒团队，在默认认为质量问题需要更重的 orchestration stack 之前，应该先测试这种低成本的 manager-worker critique loop。

Amjad Masad used three posts to hammer the same product thesis: Replit wants agent coding to feel fast, interactive, and accessible enough that users can build a lot before they hit a paywall. The OpenAI partnership post added distribution context, but the clearer signal is product positioning: lower both cost and latency so coding agents feel like a live tool rather than a metered batch job.

Amjad Masad 用三条帖子反复强调同一个产品判断：Replit 希望 agent coding 足够快、足够交互化，也足够容易上手，让用户在碰到 paywall 之前就能完成很多构建。与 OpenAI 的合作提供了分发背景，但更清晰的信号其实是产品定位：同时降低成本与延迟，让 coding agents 更像一个实时工具，而不是按量计费的 batch job。

Nikunj Kothari's argument is that as venture funds get larger, small wins stop mattering mathematically, so investors are pushed toward companies with extreme upside and are less sensitive to entry price if the outcome can be huge enough. That is less a motivational speech than a financing constraint: founders are increasingly being judged against return math shaped by Anthropic-, OpenAI-, SpaceX-, and Cursor-scale outcomes.

FPV Ventures 合伙人 Nikunj Kothari 的观点是，随着 venture funds 规模继续变大，小胜利在数学上越来越不重要，因此投资人会被迫偏向那些上行空间极端巨大的公司；只要潜在结果足够大，他们对 entry price 的敏感度也会下降。这与其说是励志发言，不如说是融资约束：founders 正越来越多地被放到 Anthropic、OpenAI、SpaceX 和 Cursor 这类级别的回报数学里衡量。

Anthropic says the recent wave of Claude Code quality complaints came from three product-layer changes rather than a model or API regression: lowering default reasoning effort from high to medium, a bug that kept clearing older reasoning from idle sessions on every subsequent turn, and a prompt change that over-compressed verbosity between tool calls and final answers. The company says all three issues were fixed by April 20 in v2.1.116, and the practical lesson is that small harness or prompt changes can look like model degradation when they silently affect memory, latency, and answer quality at once.

Anthropic 表示，最近一轮关于 Claude Code 质量下降的反馈，根源并不是 model 或 API regression，而是三个 product-layer changes：把默认 reasoning effort 从 high 降到 medium；一个会在 idle session 之后每一轮都继续清除旧 reasoning 的 bug；以及一条把 tool calls 之间与 final answer 的文字量压得过紧的 prompt change。公司称这三个问题都已在 4 月 20 日发布的 v2.1.116 中修复；更实际的教训是，细小的 harness 或 prompt 改动，也可能因为同时影响 memory、latency 与 answer quality，而在用户眼中表现成“模型退化”。

Anthropic's architecture write-up for Managed Agents is fundamentally about keeping long-running agents recoverable. It separates the durable session from the harness and the sandbox, argues that credentials should live in resource configuration or external vaults rather than inside the execution environment, and reports that provisioning sandboxes only when needed cut median time-to-first-token by about 60% and p95 by more than 90%. The broader takeaway is that agent infrastructure is starting to look like systems engineering again: stable interfaces on top, replaceable implementations underneath.

Anthropic 关于 Managed Agents 的架构文章，本质上是在讨论如何让 long-running agents 具备可恢复性。它把 durable session、harness 与 sandbox 拆开，主张把 credentials 放在 resource configuration 或外部 vault 中，而不是塞进 execution environment；同时还报告说，只在需要时才 provision sandbox，使 median time-to-first-token 下降约 60%，p95 下降超过 90%。更大的启示是，agent infrastructure 正再次变得像 systems engineering：上层接口稳定，下层实现可以替换。

Anthropic is extending Claude Managed Agents with self-hosted sandboxes in public beta and private MCP access through MCP tunnels in research preview. The concrete promise is that tool execution, files, packages, services, and private systems can stay inside customer-controlled or managed infrastructure, while Anthropic continues to run orchestration, context management, and recovery. That directly addresses the biggest enterprise objection to agent deployment: not whether the model is capable, but whether the execution boundary is acceptable.

Anthropic 正在把 Claude Managed Agents 扩展到 self-hosted sandboxes（public beta）以及通过 MCP tunnels 访问 private MCP servers（research preview）。它给出的具体承诺是：tool execution、文件、packages、services 和 private systems 可以留在客户自控或托管的基础设施内，而 Anthropic 继续负责 orchestration、context management 与 recovery。这实际上直指 enterprise 部署 agent 的最大顾虑：问题不只是模型能力够不够，而是执行边界是否可接受。

The takeaway: Max Hodak is treating neural interfaces less like one-off moonshots and more like compounding engineering systems. Hodak, the founder and CEO of Science, says Prima pairs a chip implanted under the retina with glasses that project laser images, bypassing damaged rods and cones to restore a usable visual signal; after Science acquired the underlying French company, the product reached European marketing approval in July and first sales were expected within weeks.

核心结论是：Max Hodak 正把 neural interfaces 看成一种可以持续迭代、不断复利的 engineering system，而不是一次性的科幻式豪赌。Science 的 founder and CEO Hodak 介绍说，Prima 通过植入视网膜下方的芯片，配合投射激光图像的眼镜，绕过受损的 rods and cones 来恢复可用视觉信号；Science 收购底层法国公司之后，该产品已在 7 月获得欧洲 marketing approval，并预计在数周内启动首批销售。

His most interesting claim is philosophical but grounded in product reality: "The brain very literally, very clearly, plainly is a computer." In practice, that means starting with medically useful interfaces such as vision, proving that patients can use even a limited black-and-white field of view for tasks like reading and puzzles, and then improving grayscale, color, and field of view step by step rather than waiting for a single decisive scientific breakthrough.

他最值得注意的判断，带有哲学意味，但又扎根于产品现实：“The brain very literally, very clearly, plainly is a computer.” 落到实践上，这意味着先从 vision 这类有明确医疗价值的接口切入，证明患者即使只拥有有限的黑白视野，也能完成阅读和解谜等任务；然后再一步一步提升 grayscale、color 与 field of view，而不是等待某个“一锤定音”的科学突破。