Agents can be smart and still unreliable for real work when they lack domain expertise. This talk argues that the next step isn’t more agent scaffolding—it’s packaging reusable expertise as “Skills” that agents can load when needed.

Key takeaways

• Breaks down what Skills are: organized folders of procedural knowledge (including scripts/tools) that can be versioned, shared, and composed.
• Explains why code and the file system can act as a universal interface, while Skills supply the missing domain context.
• Walks through “progressive disclosure” to protect the context window and enable libraries of hundreds or thousands of Skills.
• Maps an emerging architecture: agent loop + runtime environment + MCP servers for connectivity, with Skills providing expertise—and how that supports deployment across domains.