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

About this video

• Video Title: Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic
• Channel: AI Engineer
• Speakers: Barry Zhang, Mahesh Murag
• Duration: 00:16:22

Overview

This video discusses the evolution of AI agents, highlighting the limitations of current agent designs that possess intelligence but lack domain expertise. The speakers, Barry Zhang and Mahesh Murag from Anthropic, introduce "Skills" as a solution. Skills are presented as organized collections of files containing procedural knowledge, acting as a more effective way to equip agents with specialized capabilities for real-world tasks. The talk explores the development of skills, the growing ecosystem, and the future potential of agents learning and creating their own skills.

Key takeaways

• The "Agent Skills" Paradigm: The video proposes shifting from building standalone agents to developing "Skills," which are modular packages of procedural knowledge (files, scripts, etc.) that agents can dynamically load. This approach addresses the gap where agents have intelligence but lack specific expertise.
• Code as a Universal Interface: The speakers argue that code, rather than specific domain logic, serves as a universal interface for agents to interact with the digital world, enabling greater universality in agent design.
• Limitations of Traditional Agents: Current agents are compared to brilliant but inexpert individuals (like a "300 IQ mathematical genius" for taxes instead of an experienced professional). They lack domain expertise, can't absorb user expertise well, and don't learn over time.
• Skills as Organized Folders: Skills are described as simple, organized folders containing scripts that act as tools. They are versionable, shareable, and allow agents to execute specific tasks more consistently and efficiently.
• Progressive Disclosure for Scalability: Skills use a progressive disclosure mechanism, where only metadata is initially shown to the model, allowing for a large number of skills to be managed without overwhelming the agent's context window.
• Diverse Skill Ecosystem: The ecosystem of skills is growing rapidly, with categories including foundational skills (adding new general capabilities), third-party skills (from partners), and enterprise-specific skills (for internal best practices and bespoke software).
• Future of Skills: Future developments include skills becoming more complex (packaging executables, binaries), treating skills like software (with testing, versioning, and dependency management), and enabling agents to create their own skills, leading to continuous learning and an evolving knowledge base.