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

Anthropic dropped a revenue bomb on stage this week. Dario Amodei casually revealed that his company’s annualized growth in Q1 2026 didn’t just meet its 10x target—it blew past it to 80x. Eighty. Times. Let that sink in. This isn’t just a startup having a good quarter; this is a tectonic shift in the economics of AI infrastructure, and it explains every frantic partnership, every caching metric, and every new developer tool that came out of their “Code with Claude 2026” event.

The entire day-long showcase felt less like a product launch and more like a field report from the front lines of scaling chaos. The narrative wasn’t about a smarter model—it was about the staggering operational machinery required to support one. Take GitHub’s appearance. Their Chief Product Officer wasn’t talking about coding magic; he was talking about cache hit rates. The goal? Keep it above 94%. Drop below 70%, and your prompt engineering is broken. This is the new, unglamorous reality of AI at scale: it’s not just about getting the right answer, it’s about getting the same answer, in the same way, billions of times, for a fraction of a cent. When Rodriguez compares it to high-frequency trading, he’s not being hyperbolic. This is the new latency war.

And the solutions are getting cleverly, almost sneakily, hierarchical. Anthropic’s “advisor” pattern—where a small, cheap model like Haiku handles the grunt work and only calls in the big, expensive Opus for the genuinely hard bits—is the operationalization of a core economic principle. It’s the AI equivalent of a junior associate drafting a brief, with a partner stepping in only for the final, critical review. They’re not hiding that this is a cost play; they’re celebrating it as architectural intelligence. It’s a mature, almost corporate, approach to AI development, far removed from the “throw more GPUs at it” phase.

This pragmatism extends to the developer experience. The updates to Claude Code aren’t flashy AI demos; they’re quality-of-life fixes for the humans in the loop. Remote control so you can start a debugging session on your workstation and finish it on your phone? That’s for the engineer who is always “on.” The reworked GUI with split views and pinned messages is about managing complexity, not showcasing raw capability. It’s an admission that the interface between human and AI is now a major product frontier in itself. The most telling feature might be the “routines”—scheduled, triggered prompts via cron or webhooks. It’s the quiet evolution of Claude from a conversationalist into a background utility, a silent co-worker running tasks on a timer.

Then came the managed agents pitch, where Jess Yan and Lance Martin nailed the real bottleneck: it’s not intelligence, it’s infrastructure. This is the dirty secret of every ambitious AI startup right now. Building the model is one thing; building a safe, scalable, observable runtime for autonomous agents is a whole different, orders-of-magnitude harder problem. Sandboxed execution, checkpointing, credential scoping—this is the plumbing that makes production-grade AI possible. Anthropic is selling the picks and shovels to its own gold rush, and frankly, that might be the smarter business.

The philosophical veneer came from Daniela Amodei’s “hold light and shade” cultural value. It’s a poetic way of framing the central tension of their business: building maximally capable systems while meticulously constraining them. But in the context of 80x growth, it feels less like a gentle ethos and more like a necessary operational mantra. When you’re the utility powering a surge this massive, your safety systems can’t be an afterthought; they are the product’s core reliability feature.

So what’s the real story from San Francisco? It’s that the AI race has decisively entered its infrastructure and optimization phase. The benchmark wars are giving way to cost-per-token wars. The demos of raw intelligence are being overshadowed by demos of clever caching, model routing, and developer workflow integration. Anthropic’s explosive growth is proving the market is voracious, and their response is to build the enterprise-grade scaffolding around their model core.

This is the moment AI stops being a novelty and starts being a utility. And like all utilities, its future will be determined not by the purity of its source, but by the resilience of its grid, the efficiency of its distribution, and the ingenuity of the apps built on top of it. The 80x number isn’t just a brag; it’s a declaration that the messy, complex, and profoundly un-sexy work of scaling has begun in earnest. The real competition is no longer just who has the best brain, but who can feed it, manage it, and deploy it without the whole system catching fire. Anthropic seems to get that. For now, at least, they’re running at the front of a pack that’s still figuring out how to tie its shoes.

当Anthropic宣布Claude Code可以自主创建和销毁隔离分支时，台下开发者们的掌声背后，藏着一丝不易察觉的颤栗。这已经不是一个关于“AI辅助编程”的故事了，而是一场关于“谁在掌控代码”的微妙权力移交。旧金山Code with Claude 2026大会的直播，表面是技术秀场，内核却是一份AI原生时代新秩序的预告。

Dickson Tsai展示的Auto模式，本质上是把人类程序员最敏感的权限决策——那些关于“这行代码会不会搞崩生产环境”的终极恐惧——外包给了一个分类器。美其名曰筛查破坏性操作，但谁来审查这个分类器的决策？当AI能自行创建分支、执行计划任务（routines）时，它就不再是“助手”，而是悄然占据了开发流程中一个关键的决策节点。远程控制、手机续接这些体验优化固然讨巧，但真正值得深思的是：Anthropic在用这些看似便利的功能，悄悄重塑开发者的工作习惯，将他们的工作流更深地锚定在Claude的生态里。这不是赋能，这是柔性的平台绑定。

GitHub首席产品官Mario Rodriguez对缓存命中率的执念，暴露了大规模AI应用的残酷经济现实。94%的目标听起来像个运维指标，实则是生死线。他将之比作高频交易，这话一针见血——在AI API调用成为基础水电的时代，效率上1%的差异，就是百万美元的现金焚烧或节省。但这里有个辛辣的讽刺：当整个行业为模型的“智能”进步欢呼时，最先被极致优化的，却是最不“智能”的部分——缓存、提示词组装、调用路由。这像极了互联网早期，最先进的算法最终败给了CDN和负载均衡。

而Anthropic与GitHub演示的“advisor策略”（小模型调用大模型），则彻底撕下了“AGI”那层温情脉脉的面纱，暴露出赤裸的工程算计。用Haiku处理简单任务，仅在疑难杂症时召唤Opus，这套组合拳的潜台词是：智能是可以被精细拆分、按需采购的商品。它追求的不是通用的、全能的“神谕”，而是成本效益最优解。GitHub内部那个昵称“Rubber Duck”的批评模块，更是这种实用主义哲学的体现——AI不仅写代码，还要自己审查自己，人类开发者则退到了“评审AI评审”的奇妙位置。生产瓶颈从“智能”转向“基础设施”的论断非常准确，但更深层的问题是：当基础设施的复杂性由AI主导构建和维护时，人类工程师还有能力理解和掌控它吗？

Dario Amodei抛出的“年化收入增长80倍”是一颗深水炸弹。这个数字早已脱离常规增长曲线，进入了一个由算力、资本和紧急需求共同驱动的狂热区间。他将此与SpaceX合作缓解算力压力联系起来，恰恰点明了当前AI竞赛的实质：这已不是单纯的算法竞赛，而是演变为整合太空通信、能源供给的超级工程竞赛。所谓“掌控光明与阴暗面”的文化价值观，在如此陡峭的增长斜率面前，听起来更像是一种面对失控风险的自我警醒与公关话术。当增长本身成为最大信仰时，安全护栏能承受多少来自商业压力的撕扯？

整场大会，Anthropic都在编织一个叙事：模型能力跃迁，正在倒逼产品、组织乃至基础设施的全面重构。他们展示的Managed Agents沙箱、检查点、凭据控制，是在为AI的“自治”铺路，试图定义下一代开发者平台的标准协议。但危险也正在于此。当一家公司同时定义了模型的智能、工具的交互、甚至基础设施的规范时，它正在成为事实上的“AI时代的操作系统”供应商。开发者“是Claude最重要的用户”——Daniela Amodei的这句赞美，或许换个角度理解更真实：开发者是Claude生态得以扩张和巩固的最重要“节点”。

这场大会的真正看点，不是Claude Code又多了哪个功能，而是Anthropic如何巧妙地将“开发者工具”升级为“开发者生存环境”的构建。他们卖的不再是锤子，而是整个木匠铺的租赁权，并且这个铺子会自己生长、自己优化。未来的开发者，或许会更高效，但也可能更像一个在AI精密设计好的流水线旁，负责处理异常和进行创造性点睛的“系统维护员”。这究竟是进化还是异化，可能要等到80倍增长的热潮退去，人们才能看清沙滩上究竟留下了什么。