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

Summary

In this talk, the speakers Barry and Mahesh discuss the evolution of agent technologies, focusing on the shift from building agents to developing agent skills. They highlight how agent skills, which are organized folders of procedural knowledge, enhance agent capabilities by adding domain expertise. Through their system, agents can use skills to execute tasks efficiently, drawing on a growing ecosystem of foundational, third-party, and enterprise-specific skills. The session outlines the future of general agents, emphasizing the importance of skills in enabling agents to continually evolve and adapt, with an open invitation for developers to participate in this exciting advancement.

Takeaways

• 😀 The concept of 'agents' has evolved, and while they are capable, they still lack expertise in specific domains, necessitating the development of specialized skills.
• 😀 The new paradigm for agents emphasizes a tight coupling between models and runtime environments, with the belief that code alone is the universal interface to the digital world.
• 😀 Instead of building domain-specific agents, the focus has shifted to building skills—organized collections of files containing procedural knowledge that agents can use.
• 😀 Skills are simple and composable, allowing anyone, human or agent, to create and use them easily. They are compatible with existing systems, such as Git and Google Drive.
• 😀 Traditional tools have limitations, but code-based tools (scripts) can be self-documenting, modifiable, and stored in the file system for later use.
• 😀 Agent skills are progressively disclosed at runtime, ensuring only necessary information is shown to the agent while maintaining organizational efficiency.
• 😀 The ecosystem of agent skills has rapidly expanded, with foundational, third-party, and enterprise-specific skills emerging, allowing for enhanced agent capabilities.
• 😀 Skills are becoming more complex and include executable files, software, and other resources, leading to longer build times and more advanced integrations.
• 😀 There is growing excitement around skills being developed by non-technical users in fields like finance, legal, and recruiting, making agents more accessible and useful for everyday tasks.
• 😀 The emerging architecture for general agents involves agents with a model loop, a runtime environment, connected to external MCP servers, and equipped with a library of skills to enhance their capabilities.
• 😀 Skills enable continuous learning and memory, making agents smarter over time. They can evolve and adapt, acquiring new capabilities while discarding outdated ones, ensuring better performance with each interaction.

Q & A

What is the main shift in focus presented in the script regarding agent systems?

The main shift is from building standalone agents to focusing on creating 'skills'—modular, composable units of procedural knowledge that enhance agents' capabilities by providing them with domain-specific expertise.

Why were agents initially not sufficient for real work despite their intelligence?

While agents demonstrated intelligence and capabilities, they lacked the domain expertise necessary for executing specific tasks effectively. This led to gaps in their functionality, prompting the development of skills.

What is a skill in the context of agent systems?

A skill is an organized collection of files that package procedural knowledge and tools, like scripts, enabling agents to perform specific tasks more effectively. These skills are simple, modular, and easily shareable.

How do skills help address the issue of agents lacking expertise?

Skills provide domain-specific knowledge and tools that agents can use, which fills the gap where agents lack expertise. This allows agents to perform specialized tasks with more consistency and precision.

What was the role of 'cloud code' in this transition?

Cloud code served as a general-purpose agent, highlighting the realization that code is the universal interface for interacting with digital systems. It allowed agents to generate and organize data, analyze it, and synthesize insights through code, demonstrating scalability.

What are the different types of skills mentioned in the ecosystem?

The script identifies three main types of skills: foundational skills (general or domain-specific capabilities), third-party skills (created by ecosystem partners), and enterprise-specific skills (tailored for internal use within organizations).

How has the skill ecosystem evolved in the five weeks since its launch?

Since the launch, the skill ecosystem has rapidly expanded, with thousands of skills created. Skills are becoming more complex, with some incorporating executables, binaries, and code. Additionally, there's a growing trend of non-technical people contributing to the ecosystem.

What is the role of MCP (Model-Connectivity Protocol) servers in the new architecture?

MCP servers provide the external connectivity and tools that agents use to interact with the world. When combined with the right set of skills, MCP servers help agents execute more complex tasks, adding depth and context to their capabilities.

What future challenges and opportunities are foreseen for skills in this ecosystem?

Future challenges include developing better tooling for testing, evaluating, and versioning skills, as well as improving the predictability of agents by ensuring dependencies between skills and MCP servers are clearly defined. The goal is to make skills easier to build and integrate into agent systems.

How do skills contribute to the continuous learning of an agent like Claude?

Skills enable continuous learning by allowing agents to accumulate procedural knowledge over time. This knowledge can be transferred across agent versions, allowing an agent to become more effective and knowledgeable the longer it interacts with users.