Back to Videos

Don't Build Agents, Build Skills Instead

Speakers Barry Zhang & Mahesh Murag, Anthropic
Event AI Engineer Conference
Date December 8, 2025
Duration 16:22
Anthropic Agent Skills Architecture MCP
TL;DR

Anthropic introduces "Agent Skills" - a new paradigm for extending AI agents. Instead of building separate agents for each use case, skills are organized folders of procedural knowledge that give general-purpose agents like Claude Code domain expertise. Skills are simple (just folders with markdown and scripts), progressively disclosed to protect context windows, and already being adopted by Fortune 100 companies to teach agents their organizational best practices.

Key Takeaways

Summary

The Shift to General-Purpose Agents

The speakers explain how their thinking has evolved since launching MCP. They initially believed different domains would need entirely different agents with specialized scaffolding. Instead, they discovered that code itself is the universal interface to the digital world.

"We realized that code is not just a use case but the universal interface to the digital world."

Claude Code can generate financial reports by calling APIs, organizing data in the file system, analyzing with Python, and synthesizing insights - all through code. The core agent scaffolding becomes as thin as "just bash and file system."

The Domain Expertise Problem

While agents are brilliant, they lack domain expertise. The speakers use an analogy: Who should do your taxes - a 300 IQ genius or an experienced tax professional? The professional wins because you need consistent execution from someone who already knows the domain.

"Agents today are a lot like Mahesh. They're brilliant, but they lack expertise."

What Skills Are

Skills are organized collections of files that package composable procedural knowledge for agents. In simpler terms: they're folders with a standard structure.

Key characteristics:

"Skills are organized collections of files that package composable procedural knowledge for agents. In other words, they're folders."

The Growing Ecosystem

Five weeks after launch, thousands of skills have been created across three categories:

  1. Foundational Skills: Give agents new general capabilities (e.g., Anthropic's document skills for creating professional Office documents, Cadence's scientific research skills for bioinformatics)
  2. Third-Party Skills: Built by partners for their products (Browserbase's Stagehand for browser automation, Notion for workspace research)
  3. Enterprise Skills: Company and team-specific skills for organizational best practices and internal software usage

"Fortune 100s are using skills as a way to teach agents about their organizational best practices and the weird and unique ways that they use bespoke internal software."

The Emerging Architecture

The general agent architecture is converging on:

"MCP is providing the connection to the outside world while skills are providing the expertise."

Skills as Continuous Learning

Skills represent a concrete step toward continuous learning. When you first start using Claude, the standardized format guarantees that anything Claude writes down can be efficiently used by a future version of itself.

"Our goal is that Claude on day 30 of working with you is going to be a lot better than Claude on day one."

The Computing Analogy

The speakers compare the AI stack to traditional computing:

"A few companies build processors and operating systems, but millions of developers have built software that encoded domain expertise. We hope that skills can help us open up this layer for everyone."

Notable Quotes

"We stopped building agents and started building skills instead."

"Code is not just a use case but the universal interface to the digital world."

"Skills are organized collections of files that package composable procedural knowledge for agents. In other words, they're folders."

"It's time to stop rebuilding agents and start building skills instead."

References

Mentioned Products & Companies

Skill Types