Don't Build Agents, Build Skills Instead

Overview

Barry Zhang and Mahesh Murag from Anthropic present agent skills — a new paradigm for extending general-purpose agents like Claude Code. Instead of building separate agents for each domain, they argue we should build skills: organized folders of procedural knowledge that any agent can pick up and use.

The Problem: Intelligence Without Expertise

Agents today are like a 300 IQ genius with no domain experience. They can figure things out from first principles, but you don't want that for your taxes — you want a domain expert who executes consistently.

Current agents:

• Are brilliant but lack expertise
• Can do amazing things with enough guidance, but miss important context upfront
• Can't absorb your expertise well
• Don't learn over time

What Are Skills?

Skills are organized collections of files that package composable procedural knowledge for agents. In practice: they're folders.

This simplicity is deliberate:

• Anyone (human or agent) can create and use them
• They work with existing tools — version in Git, share via Google Drive, zip and send
• They can include scripts as tools alongside instructions

Why Files Over Traditional Tools?

Traditional tools have problems:

• Poorly written instructions, ambiguous behavior
• When the model struggles, it can't modify the tool — it's stuck
• They always live in the context window

Code/files solve these issues:

• Self-documenting
• Modifiable by the agent
• Can live in the file system until needed

Progressive Disclosure

Skills protect the context window through progressive disclosure:

1. At runtime, only metadata is shown (skill name/description)
2. When needed, the agent reads the SKILL.md with core instructions
3. Everything else (scripts, assets, examples) is organized in the folder for on-demand access

This allows fitting hundreds of skills while keeping them composable.

Types of Skills

Foundational Skills

General or domain-specific capabilities the agent didn't have before.

• Document skills (by Anthropic): create/edit professional Office documents
• Scientific research skills (by Cadence): EHR data analysis, bioinformatics libraries

Third-Party Skills

Partners building skills for their own products:

• Browserbase/Stagehand: browser automation and web navigation
• Notion: deep research over your entire workspace

Enterprise Skills

Company and team-specific skills:

• Teaching agents organizational best practices
• Developer productivity teams deploying to thousands of developers
• Encoding code style, internal tooling knowledge, and workflows

The Emerging General Agent Architecture

The architecture is converging on:

1. Agent loop — manages internal context and token flow
2. Runtime environment — file system + ability to read/write code
3. MCP servers — tools and data from the outside world (connectivity)
4. Skills library — hundreds/thousands of skills pulled into context on demand (expertise)

MCP provides the connection; skills provide the expertise.

Giving an agent new capabilities in a new domain = equipping it with the right MCP servers + the right skills.

Future Directions

Skills as Software

As skills get more complex (packaging executables, binaries, assets), they need:

• Testing and evaluation — measuring output quality
• Trigger tooling — ensuring skills load at the right time for the right task
• Versioning — tracking behavior changes over time
• Dependencies — skills referring to other skills, MCP servers, or packages

Continuous Learning

Skills are designed as a concrete step towards continuous learning:

• Anything Claude writes down can be used by a future version of itself
• Memory becomes tangible — capturing procedural knowledge for specific tasks
• Claude can acquire, evolve, and drop skills as needed
• Goal: Claude on day 30 is much better than Claude on day 1

Sharing and Distribution

The most exciting vision: a collective, evolving knowledge base curated by people and agents:

• New team members get an agent that already knows the team's practices
• Community-built skills make everyone's agents more capable
• Similar to how community MCP servers benefit all agents

The Computing Analogy

Computing	AI
Processors	Models
Operating Systems	Agent runtimes
Applications	Skills

A few companies build processors and OSes, but millions of developers build applications. Skills open up this "application layer" for everyone.

Key Takeaways

1. Code is the universal interface — Claude Code is actually a general-purpose agent, not just a coding tool
2. Skills = folders — deliberately simple so anyone can create and share them
3. Progressive disclosure protects context window while enabling hundreds of skills
4. MCP + Skills are complementary — MCP for connectivity, skills for expertise
5. Stop rebuilding agents — the agent is universal; customize it with skills instead
6. Skills enable continuous learning — agents that get better the more you use them