Descriptions:
Barry Zhang and Mahesh Murag from Anthropic introduce agent skills at AI Engineer, presenting a new paradigm for building AI systems. Rather than constructing separate specialized agents for each domain, they argue that a single general-purpose agent—like Claude Code—can handle virtually any task through code as a universal interface. The core problem they identify: agents have broad intelligence but lack domain expertise, analogous to having a brilliant generalist rather than an experienced specialist.
Agent skills are deliberately simple: organized folders containing markdown instructions, scripts, and reusable tools that package procedural knowledge for agents. They integrate with existing infrastructure—Git, Google Drive, zip files—and are progressively disclosed at runtime to protect context windows, enabling hundreds of skills to be composed simultaneously. Five weeks after launch, the ecosystem had grown to thousands of skills spanning foundational capabilities, third-party integrations, and enterprise-specific knowledge.
The talk outlines an emerging architecture where MCP servers handle external connectivity while skills supply domain expertise. Fortune 100 companies are already using skills to encode organizational best practices and internal software workflows, while non-technical employees in finance, legal, and recruiting are building skills without writing code—validating the vision of expertise-packaged agents accessible to anyone with a computer.
📺 Source: AI Engineer · Published December 08, 2025
🏷️ Format: Keynote Launch
