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

The Challenge with Current Agents

• 💡 While agents possess intelligence and capabilities, they often lack the specific domain expertise required for real-world tasks, similar to a brilliant person without professional experience.
• 🧠 Agents struggle to absorb expertise effectively, maintain important context, and learn consistently over time from interactions.

Introducing Agent Skills

• 🔑 Agent skills are defined as organized collections of files that package composable procedural knowledge for agents, essentially functioning as simple folders.
• ✅ This design is deliberate, allowing anyone (human or agent) with a computer to create and utilize them, integrating seamlessly with existing tools like Git for versioning.
• 🛠️ Unlike traditional tools with ambiguous instructions, skills use code as a self-documenting and modifiable interface, residing in the file system to protect the agent's context window.
• 🚀 Skills employ progressive disclosure, showing only metadata initially and loading full instructions (skill.md) when needed, enabling hundreds of skills to be composable.

A Rapidly Growing Ecosystem

• 📈 Since their launch, skills have fostered a quickly expanding ecosystem across various types, demonstrating their versatility and impact.
• 🎯 Foundational skills provide agents with new general or domain-specific capabilities, such as document creation or scientific research, enhancing their core functionality.
• 🤝 Third-party skills allow partners to integrate their software, enabling agents to navigate the web or understand complex workspaces more effectively.
• 🏢 Enterprise-specific skills are gaining significant traction, helping large organizations embed best practices, manage bespoke internal software, and improve developer productivity.

The Evolving Agent Architecture

• 🧩 The emerging architecture for general agents combines an agent loop (managing context), a runtime environment (file system, code execution), and MCP servers (external tools/data).
• 📚 Agents can dynamically pull from a library of hundreds or thousands of skills into context only when needed for a specific task.
• 🚀 This integrated approach allows for the rapid deployment of agents to new verticals like financial services and life sciences, making them immediately more effective for professionals.

Future Directions and Continuous Learning

• 🌱 Future development focuses on treating skills like software, including testing, evaluation, and versioning, to ensure reliability and track evolution.
• 🔗 Plans include enabling skills to explicitly depend on other skills, MCP servers, and packages, enhancing predictability and composability in diverse runtime environments.
• 💡 Skills are a concrete step towards continuous learning for agents; Claude can already create skills, ensuring that learned procedural knowledge is transferable and leads to significant improvement over time.
• 🌐 Analogous to applications on an operating system, skills are envisioned as a collective, evolving knowledge base curated by both people and agents, fostering shared capabilities across organizations and communities.