Code 2025
Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic
Overview
This talk proposes a shift from building standalone AI agents to developing modular "skills" that agents can utilize. The core argument is that while agents possess intelligence, they often lack the specific domain expertise required for real-world tasks. Skills, defined as organized collections of files containing procedural knowledge and tools, offer a more composable and expertise-driven approach to enhancing agent capabilities. This paradigm aims to make AI more accessible and adaptable by packaging expertise into shareable units.
Who should watch
- AI engineers and builders looking to enhance agent functionality.
- Product Managers seeking to integrate specialized AI capabilities into products.
- Developers struggling with agents lacking domain-specific expertise.
- Teams aiming to standardize and share organizational knowledge with AI.
- Anyone interested in the evolution of agent architecture beyond monolithic agents.
Key takeaways
- Agents often excel in general intelligence but lack the specific expertise needed for complex tasks, akin to a brilliant mathematician versus an experienced tax professional.
- Skills are proposed as organized collections of files, essentially folders containing procedural knowledge and scripts acting as tools, which agents can dynamically load.
- Code serves as a universal interface and a self-documenting, modifiable primitive for skills, overcoming limitations of traditional, ambiguous tool instructions.
- Skills are progressively disclosed to the model at runtime, with only metadata initially shown to manage context window limitations and enable composability.
- An ecosystem of skills is rapidly growing, including foundational, third-party, and enterprise-specific skills, demonstrating broad adoption and diverse applications.
- The architecture for general agents is converging on an agent loop, a runtime environment, MCP servers for external connectivity, and a library of skills for expertise.
- Skills are being treated like software, with future development focusing on testing, evaluation, versioning, and explicit dependency management to improve predictability and composability.
- The ultimate vision is a collective, evolving knowledge base of capabilities curated by people and agents, making AI more personalized and effective within organizations and the broader community.
Notable quotes
*We think code is all we need.*
*Skills are organized collections of files that package composable procedural knowledge for agents.*
*We hope that skills can help us open up this layer for everyone.*
Unofficial community note. Prefer the recording for nuance.