Anthropic 平台：从知识到执行再到协调

Anthropic Platform: From Knowledge to Execution to Coordination

凯特琳·莱西 Katelyn Lesse · 红杉资本 (Training Data) · 2026-07-14 · 约 49 分钟 · 原视频 ↗

本期速览 · Overview

Anthropic 平台团队讨论他们的北极星目标——为构建者提供工具，从知识层到执行层再到协调层的演进，以及保持内部和外部平台一致的理念。

Anthropic's platform team discusses their north star of providing tools for builders, the evolution from knowledge to execution to coordination layers, and their philosophy of keeping internal and external platforms consistent.

要点 · TL;DR

Anthropic 的平台将 AI 抽象为知识、执行和协调三层。
Anthropic's platform abstracts AI into knowledge, execution, and coordination layers.
令牌任务分配和成本优化是最大化每美元智能的关键。
Token job assignment and cost optimization are key to maximizing intelligence per dollar.
开放生态系统和 MCP 等标准实现互操作性和实验。
Open ecosystems and standards like MCP enable interoperability and experimentation.

核心观点 · Key points

平台抽象层：知识、执行、协调。
Platform abstraction layers: knowledge, execution, coordination.
Token 的任务和策略是最大化每美元智能的关键。
Token jobs and strategies are key for maximizing intelligence per dollar.
上下文工程和框架对于构建卓越的智能体至关重要。
Context engineering and harnesses are critical for building exceptional agents.
生态开放性和 MCP 等标准实现互操作性。
Ecosystem openness and standards like MCP enable interoperability.
模型路由和成本优化是 Token 最大化后的自然下一步。
Model routing and cost optimization are natural next steps after token maxing.

反共识 · Contrarian takes

框架应针对模型家族调整，而非跨模型通用。
Harnesses should be tuned to model families, not generic across models.
上下文工程被高估；高阶策略更重要。
Context engineering is overrated; higher-order strategies matter more.
智能体群只是策略的一种；Token 任务分配更深入。
Agent swarms are just one type of strategy; token job assignment is deeper.
不要限制 AI 使用；而是设计智能路由策略。
Don't cap AI usage; instead, design smart routing strategies.
形态快速演变；平台应支持实验，而非锁定。
Form factors evolve rapidly; platforms should enable experimentation, not lock-in.

本期章节 · Chapters（共 22）

协调层与平台概述 Coordination Layer and Platform Overview
平衡内外部反馈 Balancing Internal and External Feedback
抽象层次：从原语到托管代理 Layers of Abstraction: From Primitives to Managed Agents
三层抽象：知识、执行、协调 Three Layers of Abstraction: Knowledge, Execution, Coordination
知识、执行与协调层 Knowledge, Execution, and Coordination Layers
开放生态 vs 围墙花园 Open Ecosystem vs. Walled Garden
基础设施与架构理念 Infrastructure and Architecture Philosophy
演进形态与产品理念 Evolving Form Factors and Product Philosophy
Claude Tag：有主见的代理平台 TAM and Token-Heavy Verticals
框架与上下文工程最佳实践 Claude Tag as an opinionated agentic platform
令牌分配策略 Best practices for harness and context engineering
任务特定 vs 通用框架 Token allocation strategies
代理定制与控制层 Task-specific vs general harnesses
向高级用户学习 Agent Customization and Control Layers
代理模块化与连接性 Learning from Advanced Users
令牌合理化与平台策略 Agent Modularity and Connectivity
令牌最大化与成本优化策略 Token Rationalization and Platform Strategy
代理性能的第三杠杆 Token maxing and cost optimization strategies
令牌有任务与爬山法 Third lever for agent performance
企业与开发者角色 Token has a job and hill climbing
结束语 Enterprise and developer personas
Closing remarks Closing remarks

阅读全文双语转录 →