Anthropic Releases Managed Agents, Proactive Workflows, and Capability Curves at Code With Claude Anthropic在Code With Claude上发布托管式智能体、主动式工作流与能力曲线

The most explosive figure at this launch event wasn't about what new tricks Claude Code had learned, but rather the offhand remark by Anthropic CEO Dario Amodei: "Our annualized revenue in Q1 2026 grew 80x, not the planned 10x." 80 times, not 10. That number alone explains everything—why compute power suddenly became a bottleneck, why they're rushing to partner with SpaceX, and why the tone of the entire developer conference quietly shifted from "showcasing capabilities" to "how to survive and profit in this out-of-control rocket race."

Look at what they showcased. Claude Code's "Auto Mode" and "Routines" are essentially about fitting more precise reins on the increasingly clever wild beast that is AI. Auto-determining permissions, scheduled task execution—it sounds great, but it also exposes the current awkwardness of AI toolchains: intelligence itself is skyrocketing, but embedding that intelligence safely, reliably, and controllably into production workflows is a tangled mess. What Anthropic is doing feels more like designing anti-slip hoof covers and emergency braking systems for an elephant that has suddenly charged into a china shop. The feature list looks impressive, but the underlying logic is defensive.

Then there's the sharing from GitHub. CPO Mario Rodriguez talked at length about cache hit rates, stressing they need to stay above 94%, otherwise it means "flawed prompt assembly." Doesn't that sound like an F1 racer on a highway, focusing all energy on calculating braking points? While everyone discusses AI's creative potential, the platform's biggest headache is how to make hundreds of millions of API calls cheaper, even cheaper. Cache hit rate is that critical pressure point. It coldly reminds us that the current boom in AI applications is largely built on an extremely expensive infrastructure with unreasonably high efficiency demands. GitHub's "Rubber Duck" review module (a rather endearing internal nickname) underscores this: AI-written code still requires another AI module to nitpick, and the stacking of system complexity has reached dizzying heights.

Anthropic's own product manager said the bottleneck for production-grade agents is no longer intelligence, but infrastructure. Sandboxes, checkpoints, credential controls... these terms sound aesthetically technical, but translate to: we need more and safer cages, more rollback-capable save points, more permission isolation. Is this a discussion about the dawn of AGI? No, this is a serious, engineering-focused debate on "safety production responsibility assignment" in the digital world. Intelligence may be boundless, but production environments must be meticulously calculated.

Ultimately, the two founders took the stage. Daniela Amodei spoke of "the bright and dark sides," while Dario Amodei crunched numbers on revenue and compute power. The 80x hyper-growth corresponds to a strategic partnership with SpaceX to alleviate compute pressure. See, the story comes full circle: the leap in AI capability (the bright side) directly leads to exponential growth in costs and infrastructure pressure (the dark side that must be managed). Developers are the most important users? Yes, but first, they must be users who generate revenue and are willing to pay for astronomical compute costs.

The entire conference felt like a real-time earnings report for investors and a survival guide for developers. It tells you where the frontier is, but even more forcefully, it reveals how many cost-intensive pits need to be filled with real money on the road to that frontier. Claude has become more powerful and autonomous, but the cost of using it is also becoming clearer and more expensive than ever. We are being kicked out of the romantic fantasy phase of "what AI can do" and into the muddy reality of "how to run AI cheaply and safely," reeking of money and engineering details. It turns out the ticket to this revolution is a ship ticket that keeps getting more expensive.

这场发布会最炸裂的数字，不是Claude Code又学会了什么新花招，而是Anthropic CEO Dario Amodei轻描淡写抛出的那句：“我们2026年第一季度的年化收入增长了80倍，而不是计划中的10倍。” 80倍，不是10倍。这个数字本身就解释了一切——解释了为什么算力突然成为瓶颈，解释了为什么他们急着和SpaceX谈合作，也解释了为什么整个开发者大会的基调，从“展示能力”悄然转向了“如何在这场失控的火箭竞赛中存活下来并赚到钱”。

看看他们展示的东西。Claude Code的“Auto模式”和“Routines”，本质上是在给AI这头越来越聪明的野兽套上更精密的缰绳。自动判断权限、定时执行任务，听起来很美，但这恰恰暴露了当前AI工具链的尴尬：智能本身在飙升，但让智能安全、可靠、可控地嵌入生产流程，却是一团乱麻。Anthropic在做的，更像是在为一头突然冲进瓷器店的大象设计一套防滑蹄套和急停装置。功能列表很漂亮，但底层逻辑是防御性的。

然后是GitHub那边的分享。首席产品官Mario Rodriguez大谈缓存命中率，说要维持在94%以上，否则就是“提示词组装有缺陷”。这听起来像不像一个在高速公路上开F1赛车的车手，把全部精力都放在计算刹车点上？当所有人都在谈论AI的创造潜能时，平台方最头疼的却是如何让数亿次的API调用便宜一点、再便宜一点。缓存命中率就是那个命门。它冰冷地提醒我们，当前AI应用的繁荣，很大程度上是建立在一个极其昂贵、且效率要求高到变态的基础设施之上的。GitHub的“Rubber Duck”审查模块（这个内部昵称倒是亲切），也佐证了这一点：AI写的代码，还得另一个AI模块来挑刺，整个系统复杂度的堆叠已经到了令人眼晕的地步。

Anthropic自己的产品经理说，生产级智能体的瓶颈不再是智能，而是基础设施。沙箱、检查点、凭据控制……这些词汇听起来充满技术美感，但翻译过来就是：我们需要更多更安全的笼子、更多可回滚的存档点、更多权限隔离。这哪是在讨论AGI的黎明？这分明是在进行一场严肃的、关于数字世界里“安全生产责任到人”的工程化大讨论。智能可以天马行空，但生产环境必须锱铢必较。

最终，两位创始人登台，Daniela Amodei谈“光明与阴暗面”，Dario Amodei算钱和算力账。80倍的疯狂增长，对应的是与SpaceX缓解算力压力的战略合作。看，故事闭环了：AI能力的跃迁（这是光明），直接导致成本和基础设施压力的指数级增长（这是必须管控的阴暗面）。开发者是最重要的用户？没错，但首先得是能带来收入、并且愿意为天价算力买单的用户。

整场大会，像是一份给投资人的实时财报，也是一张给开发者的生存指南。它告诉你前沿在哪里，但更用力地告诉你，去往前沿的道路上，埋着多少需要真金白银去填平的坑。Claude变得更强大、更自主了，但使用它的代价，也正变得前所未有的清晰和昂贵。我们正从“AI能做什么”的浪漫幻想期，被一脚踹进“AI如何便宜且安全地跑起来”的、充满铜臭味和工程细节的泥泞现实里。这场革命的门票，原来是一张永远在涨价的船票。