Summary
Highlights
The speakers introduce the concept of 'agent skills' as a solution to a gap in current agent technology: while agents possess intelligence and capabilities, they often lack the specialized expertise needed for real-world tasks. They highlight that the paradigm for agents is moving towards a tighter coupling between the model and its runtime environment, emphasizing that 'code is all we need'.
Initially, agents were thought to be domain-specific. However, the realization that code serves as a universal interface to the digital world led to the development of general-purpose agents like Cloud Code. Despite their brilliance, these agents, like a mathematical genius without tax experience, lack domain expertise, leading to inconsistent execution and the need for contextual knowledge.
Agent skills are introduced as organized collections of files that package composable procedural knowledge for agents. These are essentially folders and are designed for simplicity, allowing anyone to create and use them. Skills differ from traditional tools by being self-documenting, modifiable, and progressively disclosed to manage context windows efficiently. An example given is Claude saving a Python script for styling slides as a skill for future use.
Since their launch, agent skills have fostered a rapidly growing ecosystem, categorized into foundational, third-party, and enterprise-specific skills. Foundational skills add general or domain-specific capabilities, third-party skills integrate with existing software (e.g., Browserbase, Notion), and enterprise skills codify organizational best practices and bespoke internal software knowledge.
The ecosystem shows trends of increasing skill complexity, with skills now packaging software, executables, and more. Skills are also complementing existing MCP servers by orchestrating workflows, providing expertise while MCP handles external connectivity. Notably, non-technical users in various functions are building skills, validating the idea that skills make agents more accessible.
The architecture for general agents is converging on an agent loop coupled with a runtime environment, connected to MCP servers for external tools and data, and enhanced by a library of skills. This allows agents to pull relevant skills only at runtime. This modular approach is helping Anthropic deploy Claude to new verticals like financial services and life sciences, equipped with domain-specific MCP servers and skills.
Future development of skills will focus on testing, evaluation, better tooling for skill management, versioning to track evolution, and explicit dependencies between skills, MCP servers, and packages. These improvements aim to make skills easier to build, integrate, and ensure predictable behavior in different runtime environments.
A key value of skills lies in sharing and distribution, fostering a collective and evolving knowledge base. Skills provide procedural knowledge, and user feedback continually refines agent capabilities. This means new team members can quickly leverage an agent that already understands team-specific practices. Claude can also create its own skills, enabling continuous learning and adaptability to changing information.
The agent stack is compared to traditional computing: models are processors, agent runtimes are operating systems, and skills are applications. Just as applications made processors and operating systems valuable by encoding domain expertise, skills aim to open up this layer for everyone to solve concrete problems. The presentation concludes by urging a shift from endlessly rebuilding agents to building skills instead.