Anthropic 公司 Cloud Code 的创建者 Boris Cherny 在 Mad Podcast 中与 Matt Turk 分享了 Cloud Code 意外成功的经历。Cloud Code 是一种在终端中运行的 Agentic 编码 AI，其受欢迎程度呈爆炸式增长，据报道，在推出仅仅五个月后，年化收入就达到了 4 亿美元。 Cherny 解释说，Cloud Code 源于个人需求，最初是一个自动化笔记和音乐控制的工具。他在试验 Anthropic 的 API 时，赋予了模型访问 bash 工具的权限，结果它自发地开始编码。这种出乎意料的智能编辑代码的能力标志着一个转折点。 Cloud Code 是一种 Agentic 编码工具，代表了软件工程领域的重大转变。传统上，工程师们在集成开发环境（IDE）中直接操作代码，类似于使用文字处理器编写代码。而 Agentic 编码颠覆了这一模式。工程师们只需描述对 AI 的期望修改，模型就会处理实际的文本编辑工作。 该工具使用命令行界面 (CLI)，因为它最初的形式最简单。它的通用性使其能够在各种操作系统和 IDE 中运行，从而适应不同的工程偏好。虽然最初只是一个基本的 CLI，但 Cloud Code 现在也作为 IDE 扩展和 GitHub Actions 提供。该公司强调构建一个最小化的界面，以展示模型适应其演变的原始能力。 一个核心原则是，产品跟随模型，而不是反过来。Cloud Code 可以使用各种模型，包括 Sonnet4、Opus4 和 Hiku。Anthropic 的模型在编码方面表现出色，这归功于公司的研究重点以及对“编码对于实现高级 AI 至关重要”的信念。这使得模型能够与世界互动并对其产生影响。 Cloud Code 专为专业软件工程师设计，可提高他们的生产力。它也可以用于非编码任务，例如数据科学家编写查询或设计师进行原型设计。Cloud Code SDK 允许用户构建超出编码范围的 Agent 应用程序。 Agentic 功能使模型能够规划、执行和迭代任务直至完成。Cloud Code 几乎可以执行人类可以执行的任何操作，包括读取和写入文件、运行命令和访问互联网，但对于潜在的危险操作，需要获得人类的批准。 Cloud Code 可以使用模型上下文协议 (MCP) 连接到公司现有的代码知识库。它通过连接到任何可用的工具来促进工具的使用。最近发布的子代理允许用户创建特定角色的代理，例如 QA 工程师或项目经理，每个代理都有定制的提示和工具。此功能可以分解工作并利用模型的功能。然而，随着模型变得更加先进，对这种刚性角色的需求可能会减少。 Cloud Code 避免使用基于代码的索引来保护隐私。相反，它使用 Agent 搜索，像人类一样执行文件搜索，并根据结果改进搜索词。Cloud .MD 文件还可以作为工程团队的共享记忆库，让每个人都能从 Agent 的学习经验中受益。 默认行为始终是人在回路中。虽然像读取文件这样的操作可以安全地自动化，但像编辑文件这样的操作需要人工批准。敏感代码可以通过使用 Anthropic API 并与银行批准的代码编辑器集成来处理。 Cloud Code 的 UI/UX 设计注重细节，使其具有视觉吸引力。鼓励工程师专注于简洁性，并使用 Quad code 来研究、回答问题、编写代码、调试或与 Git 源代码控制交互。Anthropic 的入职流程已经缩短。工程师们只需向 Quad code 提问，几天内就能开始高效工作。 Cloud Code 按 token 收费，提供 pro 和 max 订阅以及可选的 API 密钥使用。它有不同的应用场景。您可以使用它来规划项目、管理任务、编写代码、测试代码或调试生产问题。 AI 编码领域的未来是广阔的，有容纳多个参与者的空间。重点应该放在将 AI 编码工具推广到早期采用者之外的更广泛的市场。Cherny 建议开发者关注六个月后模型的能力，而不是它们当前的局限性。

Boris Cherny, the creator of Cloud Code at Anthropic, shares the story of its accidental success on the Mad Podcast with Matt Turk. Cloud Code, an agentic coding AI operating in the terminal, has exploded in popularity, reportedly generating $400 million in annualized revenue just five months after its launch. Cherny explains that Cloud Code was born out of personal need, initially a tool to automate note-taking and music control. He was experimenting with Anthropic's API and gave the model access to a bash tool and it spontaneously began coding. This unexpected ability to edit code intelligently marked a turning point. Cloud Code is an agentic coding tool that represents a significant shift in software engineering. Traditionally, engineers directly manipulate code within Integrated Development Environments (IDEs), similar to using a word processor for code. Agentic coding flips this model. Engineers describe the desired changes to the AI, and the model handles the actual text editing. The tool utilizes a command-line interface (CLI) because it was the simplest initial form factor. Its universality allows it to function across various operating systems and IDEs, accommodating diverse engineering preferences. While initially a basic CLI, Cloud Code is now also available as IDE extensions and GitHub actions. The company emphasizes building a minimal interface that showcases the model's raw capabilities to adapt rapidly to its evolution. A core principle is that the product follows the model, rather than the other way around. Cloud Code can utilize various models, including Sonnet4, Opus4, and Hiku. Anthropic's models excel at coding due to the company's research focus and a belief that coding is crucial for achieving advanced AI. It allows models to interact with and affect the world. Cloud Code is designed for professional software engineers, multiplying their productivity. It can be used for non-coding tasks as well, such as query writing by data scientists or prototyping by designers. The Cloud Code SDK allows users to build agent applications beyond coding. Agentic functionality enables the model to plan, execute, and iterate on tasks until completion. Cloud Code can perform almost any action a human can, including reading and writing files, running commands, and accessing the internet, but with human approval for potentially dangerous actions. Cloud Code can connect to a company's existing code knowledge using the Model Context Protocol (MCP). It facilitates tool use by connecting with any tool available. Recently announced sub-agents allow users to create role-specific agents like QA engineers or project managers, each with customized prompts and tools. This feature divides work and leverages the model's capabilities. However, the need for such rigid roles may diminish as models become more advanced. Cloud Code avoids code-based indexing for privacy. Instead, it uses agent search, performing file searches like a human would, refining search terms based on results. Cloud .MD files also serve as shared memory banks for engineering teams, allowing everyone to benefit from the agent's learned experiences. The default behavior is always human in the loop. While actions like reading a file can be safely automated, actions like editing a file need human approval. Sensitive code can be handled by using the Anthropic API and integrating with a bank-approved code editor. Cloud Code's UI/UX is designed with careful attention to detail, making it visually appealing. Engineers are encouraged to focus on simplicity and to use Quad code to research, answer questions, write code, debug, or interact with Git source control. The onboarding process at Anthropic has been shortened. Engineers can become productive within days by simply asking Quad code. Cloud Code charges per token, with pro and max subscriptions and optional API key usage. It has different applications. You can use it to plan projects, manage tasks, write code, test code, or debug production issues. The future of the AI coding space is vast, with room for multiple players. Focus should be on bringing AI coding tools to the broader market beyond early adopters. Cherny advises builders to focus on the capabilities of models six months from now, rather than their current limitations.