以下是内容的中文翻译：
Anthropic 的平台团队由 Caitlin 和 Angela 领导，其任务是构建他们所描述的“全球最重要的开发者平台之一”。他们的工作既涵盖面向开发者的外部 API，也包括 Anthropic 内部产品团队所使用的基础设施。
该团队以双重目标为指引。对内，他们的目标是为 Anthropic 团队提供最大助力和速度，以便快速可靠地推出“AGI 赋能产品”。对外，他们旨在为 *任何* 构建者提供工具，使其能够利用 Claude 的智能来构建自己的应用程序和系统。这包括提供原语、API、更高层次的抽象，并为 MCP 和 Skills 等行业标准做出贡献，所有这些都旨在使过去在经济上不可行的定制软件成为可能。
指导他们平台开发的核心理念是：AI 的形态（form factors）正在不断演变。他们已经从“万物皆聊天”发展到“智能体”，并预计未来还会发生更多转变。为了实现这种快速演进和“实验的民主化”，他们专注于构建一个强大的平台，提供所有构建者都能使用的统一原语，避免过度偏重内部或外部需求。
他们将平台的抽象层描述为“分层蛋糕”：
1. **知识层 (Knowledge Layer)：** 这一基础层侧重于通过 Messages API 等接口，展示 Claude 固有的设计、能力和参数。它包括标准化工具、技能和记忆，让用户理解“如何真正地利用 Claude 做事”。
2. **执行层 (Execution Layer)：** 这一层超越了纯粹的知识，专注于使 Claude 能够“执行工作”和完成任务。它涉及底层支撑系统和托管基础设施（如 Claude 托管智能体），负责处理诸如创建安全沙盒、管理会话存储以及优化上下文窗口和提示缓存等复杂问题。
3. **协调层 (Coordination Layer)：** 这是最高级也是最具未来感的层，团队目前正将精力集中于此。在这里，“策略”的概念应运而生。认识到“令牌并非真正同质化”，这一层允许组合精心设计的策略，其中令牌被分配不同的任务——例如，一个令牌负责提供建议，一个负责执行，还有一个负责反思或“设想”。这种“元协调机制”位于执行层之上，确保所有层级协同工作。路线图表明，协调层将提供越来越抽象的功能。
Anthropic 致力于构建开放的生态系统，支持广泛创新。尽管他们会构建第一方产品来探索新的形态或解决重大的市场机遇（例如 Claude Design、Claude Tag），但他们并不拘泥于解决方案的运行 *位置*。他们强调可靠性和可扩展性的架构原则，允许用户整合其自身基础设施（例如自托管沙盒、MCP 隧道）。他们的产品开发目标是“令牌密集型”领域，在这些领域中，持续迭代和生成是关键，例如编码、金融和法律，而不是一次性任务。
来自高级用户的见解极大地影响着他们的平台开发。他们观察到“AI 原生”公司在上下文工程以及为内部智能体工具主动收集信息方面进行了大量创新。他们还看到了利用计算机视觉和自动化技术，将 AI 与缺乏 API 的“传统软件”集成方面的突破。一个特别令人兴奋的进展是，客户在其自定义智能体上公开 MCP 服务器，从而使不同的 AI 系统能够作为工具进行互操作。
关于“令牌合理化”，他们警告不要简单地限制 AI 使用量。相反，他们鼓励一种战略性方法：设计“路由器”来评估任务复杂性，并将其路由到最合适的模型（对于复杂任务使用更大、能力更强的模型；对于简单任务使用成本更低的模型）。这种方法结合优化策略，旨在在管理成本的同时保持创新，就像管理 AWS 账单一样。
展望未来，团队最兴奋的是在协调层构建能力，使用户能够组合这些先进的“策略”。这包括使复杂的、多步骤的流程（如用于“错误查找”的“N 选一”方法）更易于访问和部署。重点在于定义“令牌的工作”，并提供一个模块化、企业级就绪且具有出色开发者体验的平台，既服务于大型组织也服务于个人构建者。
Anthropic's platform team, led by Caitlin and Angela, is tasked with building what they describe as "one of the most important developer platforms in the world." Their work encompasses both external-facing APIs for developers and internal infrastructure for Anthropic's own product teams.
The team operates with a dual North Star. Internally, their goal is to provide maximum leverage and speed for Anthropic teams to ship "AGI-pilled products" quickly and reliably. Externally, they aim to equip *any* builder with the tools to harness Claude's intelligence for their applications and systems. This involves offering primitives, APIs, higher-order abstractions, and contributing to industry standards like MCP and Skills, all with the vision of making economically impossible custom software achievable.
A core philosophy guiding their platform development is the belief that AI form factors are constantly evolving. They've moved from "everything's chat" to "agents," anticipating further shifts. To enable this rapid evolution and "democratization of experimentation," they focus on building a robust platform with consistent primitives available to all builders, avoiding over-indexing on either internal or external needs.
They describe the platform's abstraction layers as a "layer cake":
1. **Knowledge Layer:** This foundational layer focuses on exposing Claude's inherent design, capabilities, and parameters through interfaces like the Messages API. It includes standardizing tools, skills, and memory, allowing users to understand "how to actually do something with Claude."
2. **Execution Layer:** Moving beyond mere knowledge, this layer focuses on enabling Claude to "execute work" and perform tasks. It involves low-level harnesses and managed infrastructure (like Claude Managed Agents) that handle complexities such as spawning secure sandboxes, managing session storage, and optimizing context windows and prompt caching.
3. **Coordination Layer:** This is the highest and most futuristic layer, where the team is currently concentrating its efforts. Here, the concept of "strategies" emerges. Recognizing that "tokens aren't really fungible," this layer allows for composing orchestrated strategies where tokens are assigned different jobs—e.g., one token advises, another executes, another reflects or dreams. This "meta harness" sits atop the execution layer, ensuring everything ladders together. The roadmap points towards increasingly abstract offerings at this coordination layer.
Anthropic aims for an open ecosystem, supporting widespread innovation. While they build first-party products to explore new form factors or address significant market opportunities (e.g., Claude Design, Claude Tag), they are not precious about *where* solutions run. They emphasize architectural principles for reliability and scalability, allowing users to integrate their infrastructure (e.g., self-hosted sandboxes, MCP tunnels). Their product development targets "token-heavy" areas where continuous iteration and generation are key, like coding, finance, and legal, rather than one-off tasks.
Insights from advanced users significantly influence their platform development. They've observed "AI-native" companies innovating heavily in context engineering and proactive information gathering for internal agentic tools. They've also seen breakthroughs in integrating AI with "old-school software" lacking APIs, using computer vision and automation. A particularly exciting development is customers exposing MCP servers on their custom agents, allowing different AI systems to interoperate as tools.
Regarding "token rationalization," they caution against simply capping AI usage. Instead, they encourage a strategic approach: designing "routers" that assess task complexity and route it to the most appropriate model (larger, more capable models for hard tasks; cheaper models for simpler ones). This approach, combined with optimized strategies, aims to maintain innovation while managing costs, much like managing an AWS bill.
Looking ahead, the team is most excited about building capabilities at the coordination layer, enabling users to compose these advanced "strategies." This includes making complex multi-step processes (like "best-of-N" approaches for bug hunting) more accessible and deployable. The focus is on defining "jobs for tokens" and providing a modular, enterprise-ready platform with excellent developer experience, catering to both large organizations and individual builders.