Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic

• Anthropic is shifting focus from building specialized agents to building reusable 'Skills' that package procedural knowledge.
• Agents have high intelligence but often lack the domain expertise required for real-world tasks like tax preparation or scientific research.
• Code is viewed as the universal interface to the digital world, allowing a thin agent scaffolding (bash and file system) to handle diverse use cases.
• Agent Skills are organized folders containing composable procedural knowledge, including scripts as tools, markdown instructions, and assets.
• Skills solve issues with traditional tools by being self-documenting, modifiable, and stored in the file system rather than consuming context window space.
• Skills are progressively disclosed: only metadata is shown to the model at runtime, with full details loaded only when needed.
• The Skills ecosystem includes foundational skills (e.g., document creation), third-party partner skills (e.g., Browserbase, Notion), and enterprise-specific skills.
• Enterprise teams use Skills to encode organizational best practices, code styles, and workflows for internal software.
• Skills complement MCP servers; MCP provides connectivity while Skills provide the expertise and workflow orchestration.
• Non-technical users in finance, legal, and recruiting are creating Skills, making agents more accessible for daily work.
• Future evolution of Skills includes testing, evaluation, versioning, and explicit dependencies between Skills and packages.
• Skills enable continuous learning, allowing agents to create and refine their own capabilities over time based on experience.
• The long-term vision is a collective, evolving knowledge base where Skills created by one user benefit others in the community.
• Anthropic compares Models to processors, Agent Runtimes to operating systems, and Skills to applications that deliver value.

This page was made by pasting a link into LinkDigest, which turns a Xiaohongshu, Douyin, YouTube, TikTok or X post into text an AI agent can read — transcript, on-screen text and image descriptions.