created: 2026-04-24
updated: 2026-04-24
tags: [source, video, youtube, agent-skills, anthropic]
type: source
url: https://www.youtube.com/watch?v=CEvIs9y1uog
author: "AI Engineer" (Barry Zhang, Mahesh Murag)
published: 2025-12-09
Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic¶
Summary¶
Anthropic engineers Barry Zhang and Mahesh Murag present Agent Skills — a minimal form factor for packaging procedural knowledge that agents can dynamically load. The talk covers why Anthropic believes skills are the solution to agents lacking domain expertise.
Key Takeaways¶
The Problem: Intelligence ≠ Expertise¶
Analogy: Who do you want doing your taxes? - Mahesh — 300 IQ mathematical genius who's never done taxes - Barry — experienced tax professional who's done thousands of returns
Agents today are like Mahesh: brilliant generalists who lack domain-specific expertise. They can figure things out from first principles but don't come preloaded with specific knowledge, workflows, or institutional memory.
The Old Assumption vs. Reality¶
Old assumption: Agents in different domains will look very different — separate agent for each use case.
Reality: The agent underneath is more universal than expected. Code is the universal interface to the digital world. Claude Code is actually a general-purpose agent that can generate financial reports, analyze data in Python, synthesize insights — all through bash and filesystem.
What Are Skills?¶
Skills are organized collections of files that package composable procedural knowledge for agents. In other words: they're folders.
- A
SKILL.mdmarkdown file with instructions - Scripts as tools (Python, bash, etc.)
- Reference documents, templates, examples
- Versionable in Git, shareable via Google Drive, zippable for teams
Progressive Disclosure¶
Skills are progressively disclosed to protect the context window:
1. At runtime, only metadata is shown (skill name + short description)
2. When agent decides it needs a skill, it reads the full SKILL.md
3. The rest of the folder is organized for ease of access
4. This allows hundreds or thousands of skills without context overload
Three Types of Skills¶
| Type | Description | Examples |
|---|---|---|
| Foundational | New general/domain capabilities | Document skills (Office docs), Cadence scientific research skills |
| Third-party | Partners' product-specific skills | Browserbase (Stagehand web automation), Notion workspace skills |
| Enterprise | Company/team-specific skills | Fortune 100 org best practices, internal software workflows, developer productivity team code style guides |
Skills + MCP = Complete Agent Architecture¶
| Component | Role | Analogy |
|---|---|---|
| Model | Processing power | CPU |
| Agent runtime | Orchestrates context, tokens in/out | Operating System |
| MCP servers | Connection to outside world | Network/IO |
| Skills | Expertise and procedural knowledge | Applications |
"MCP is the hands. Skills are the experience."
Emerging Agent Architecture¶
Agent Loop (context management)
↓
Runtime Environment (filesystem, code read/write)
↓
MCP Servers (external tools and data)
↓
Library of Skills (dynamically loaded on demand)
This pattern is already helping Anthropic deploy Claude to new verticals — financial services and life sciences launched immediately after skills launch, each with specific MCP servers + skill sets.
Future Directions¶
- Treat skills like software — testing, evaluation, quality measurement
- Versioning — track skill evolution and resulting agent behavior lineage
- Dependencies — skills that explicitly depend on other skills, MCP servers, packages
- Collective knowledge base — skills shared across organizations and community
- Agent-created skills — Claude already creates skills via "Skill Creator"; continuous learning as agents write their own skills from experience
The Computing Analogy¶
| Computing Layer | AI Equivalent |
|---|---|
| Processors | Models (massive investment, immense potential) |
| Operating System | Agent runtime (orchestrates processes, resources, data) |
| Applications | Skills (encode domain expertise, unique points of view) |
"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."