Anthropic 平台负责人讨论 Claude 日益增长的自主性如何推动平台从简单 API 演进到托管代理，以及自我优化模型的愿景。
Anthropic's platform leads discuss how Claude's growing autonomy drives platform evolution from simple APIs to managed agents, and the vision for self-optimizing models.
要点 · TL;DR
平台从端点演变为托管代理以获得最佳结果。 Platforms evolve from endpoints to managed agents for best outcomes.
基础设施扩展是代理产品化中最困难的部分。 Infrastructure scaling is the hardest part of productionizing agents.
未来平台将让 Claude 自我优化并生成子代理。 Future platforms will let Claude self-optimize and spin sub-agents.
核心观点 · Key points
平台演进从简单端点转向更高层次的抽象以获取最佳结果。 Platform evolution moves from simple endpoints to higher-order abstractions for best outcomes.
基础设施扩展是代理产品化中最困难的部分。 Infrastructure scaling is the hardest part of productionizing agents.
模型与框架紧密配对；通用热插拔效果减弱。 Model and harness become tightly paired; generic hot-swapping loses effectiveness.
代理需要人类参与审核和持续改进。 Agents need human-in-the-loop for review and continuous improvement.
代理成功通过可验证结果和预算衡量，而不仅仅是评估。 Success of agents measured by verifiable outcomes and budget, not just evals.
未来平台旨在让 Claude 自我优化，选择模型并启动子代理。 Future platform aims for Claude to self-optimize, selecting models and spinning sub-agents.
反共识 · Contrarian takes
框架工程对代理性能的影响比模型选择更大。 Harness engineering matters more than model selection for agent performance.
原语的路径依赖可能将模型锁定在特定优势上。 Path dependency in primitives can lock models into specific strengths.
代理应由用户而非仅工程团队拥有，以保持长期活力。 Agents should be owned by users, not just engineering teams, for longevity.
多代理编排能在多个抽象层实现性能提升。 Multi-agent orchestration enables hill climbing at multiple abstraction layers.
代理生命周期管理需要自动升级和退役流程。 Agent lifecycle management requires automated upgrades and retirement processes.
平台必须扩展以处理持续运行和自重建的代理。 Platform must scale to handle agents that constantly run and recreate themselves.
本期章节 · Chapters（共 15）
欢迎与介绍Welcome and introduction
从补全端点到托管代理的平台演进Platform evolution from completion endpoints to managed agents
Claude托管代理的当前原语Current primitives in Claude managed agents
时间通缩与内部构建代理Time deflation and building agents internally
未来愿景：Claude自我理解与扩展Future vision: Claude understanding itself and scaling
在平台上构建代理Building agents on the platform
模型框架的演进Evolution of model harnesses
为谁设计云托管代理Designing Cloud Managed Agents for Whom
内部代理用例与模式Internal agent use cases and patterns
技能与代理：人机协作的区别Skills vs. Agents: The Human-in-the-Loop Distinction
多代理编排与用例Multi-agent orchestration and use cases
衡量代理成功Measuring agent success
管理代理生命周期与退役Managing agent lifecycle and retirement
平台未来一年展望Future of the platform in one year
Claude的自我理解与代理编排Claude's self-understanding and agent orchestration