Why Anthropic Says "Don't Build Agents, Build Skills Instead"
After building Claude Code (their production coding agent), Anthropic realized something fundamental: we've been building AI agents wrong. The future isn't domain-specific agents—it's composable skills that any general-purpose agent can pull in at runtime.
Put simply, we think code is all we need.
— Barry Zhang, Anthropic (00:01:20)
Visionary talk from AI Engineer 2024
SkillsModular knowledge folders
To enterprise adoption
The Core Problem: Intelligence ≠ Expertise
Who do you want doing your taxes—Mahesh, a 300 IQ mathematical genius, or Barry, an experienced tax professional? Most people would pick Barry every time.
Who do you want doing your taxes? Is it going to be Mahesh, the 300 IQ mathematical genius, or is it Barry, an experienced tax professional, right? I would pick Barry every time. I don't want Mahesh to figure out the 2025 tax code from first principles.
— Barry Zhang, Anthropic
Watch the tax analogy (00:02:18)Current Agents Are Like Mahesh
They're brilliant (high intelligence) but lack domain expertise. They can't absorb expertise well and don't learn over time.
The Missing Piece
We've been trying to make agents smarter. The real solution is giving them domain expertise in a way they can actually use.
The Solution: Skills as Modular Knowledge Units
Instead of building a "tax agent," "coding agent," or "research agent," Anthropic proposes building one general-purpose agent and giving it skills—modular, composable folders of expertise.
Skills are organized collections of files that package composable procedural knowledge for agents. In other words, they're folders. This simplicity is deliberate.
— Barry Zhang, Anthropic
Watch (00:03:02)skill.md
Metadata and core instructions for what the skill does
Scripts/Executables
Tools and utilities the agent can use
Code/Assets
Supporting materials and context
Why This Works
Code is self-documenting, modifiable, and can live in the file system until needed. When Claude kept writing the same Python script to apply styling to slides, they saved it as a skill. Now Claude just runs the script instead of regenerating it every time.
Technical Breakthrough: Progressive Disclosure
If you have thousands of skills, you can't load them all into the context window at once. You'd run out of tokens immediately. The solution: progressive disclosure.
Step 1: Metadata Only
At runtime, only skill metadata is shown to the model to indicate what's available.
Step 2: Load on Demand
When an agent needs to use a skill, it reads the full skill.md with core instructions.
That's why skills are progressively disclosed. At runtime, only this metadata is shown to the model just to indicate that he has the skill. When an agent needs to use a skill, it can read in the rest of the skill.md, which contains the core instruction and directory for the rest of the folder.
— Barry Zhang, Anthropic
Watch (00:04:27)Real-World Adoption: Enterprise in 5 Weeks
Since launching just 5 weeks before this talk, the skills ecosystem has grown rapidly—with thousands of skills and strong enterprise adoption.
Since our launch five weeks ago, this very simple design has translated into a very quickly growing ecosystem of thousands of skills.
— Barry Zhang, Anthropic (00:05:04)
Fortune 100 Adoption
Using skills to teach agents about organizational best practices and bespoke internal software.
"We've been talking to Fortune 100s that are using skills as a way to teach agents about their organizational best practices and the weird and unique ways that they use this bespoke internal software."
— Mahesh Murag (00:06:43)
Developer Productivity
Teams serving thousands of developers are using skills to teach agents about code style best practices.
"These are teams serving thousands or even tens of thousands of developers in an organization that are using skills as a way to deploy agents like Claude Code and teach them about code style best practices."
— Mahesh Murag (00:07:03)
Why This Works
Skills solve the "bespoke software" problem. Every company has weird internal tools and processes. Skills let you teach agents about your unique environment without rebuilding the agent itself.
Third-Party Skills: The Partner Ecosystem
Partners aren't building agents—they're building skills that extend what Claude can do. This is the composable future Anthropic is betting on.
Notion
Workspace research and analysis skills that help Claude better understand your Notion workspace and do deep research.
"Notion launched a bunch of skills that help Claude better understand your Notion workspace and do deep research over your entire workspace."
— Barry Zhang (00:06:21)
Web Automation
Built a skill for their open-source browser automation tooling, Stage Hand.
"Browserbase is a pretty good example of this. They built a skill for their open-source browser automation tooling, stage hand."
— Barry Zhang (00:06:04)
Scientific Research
Scientific research skills for EHR data analysis and Python bioinformatics libraries.
"We're also really excited to see people like Cadence build scientific research skills that give Claude new capabilities like EHR data analysis and using common Python bioinformatics libraries better than it could before."
— Barry Zhang (00:05:45)
Three Emerging Trends in Skills Development
1. Skills Getting More Complex
The most basic skill is a markdown file with prompts. But we're seeing skills that package software, executables, binaries, files, code, scripts, and assets—much like traditional software applications.
"But we're starting to see skills that package software, executables, binaries, files, code, scripts, assets, and a lot more. But we think that increasingly much like a lot of the software we use today, these skills might take weeks or months to build and be maintained."
— Mahesh Murag (00:07:46)
2. Skills Complementing MCP
MCP (Model Context Protocol) provides connectivity to the outside world. Skills provide expertise and orchestration. Developers are using skills to orchestrate workflows of multiple MCP tools.
"Developers are using and building skills that orchestrate workflows of multiple MCP tools stitched together to do more complex things with external data and connectivity. In these cases, MCP is providing the connection to the outside world while skills are providing the expertise."
— Mahesh Murag (00:08:12)
3. Non-Technical Builders (Most Exciting)
People in finance, recruiting, accounting, and legal are building skills—without coding. This democratizes AI development by letting domain experts encode their knowledge directly.
"And finally, and I think most excitingly for me personally, is we're seeing skills that are being built by people that aren't technical. These are people in functions like finance, recruiting, accounting, legal, and a lot more. I think this is pretty early validation of our initial idea that skills help people that aren't doing coding work extend these general agents."
— Mahesh Murag (00:08:40)
MCP + Skills: The Perfect Pair
The division of labor is clear: MCP provides connectivity (APIs, databases, services) while skills provide expertise (how to use those connections effectively).
MCP = Connection
Connects agents to the outside world—APIs, databases, services, data sources.
Skills = Expertise
Provides orchestration and domain logic—how to use those connections effectively.
Example: Financial Reporting Skill
A "financial reporting" skill might orchestrate multiple MCP servers:
- •MCP Server for Bloomberg API (market data)
- •MCP Server for Salesforce (CRM data)
- •MCP Server for PostgreSQL (historical data)
- •Skill contains the logic for combining these sources
Clean separation: MCP servers stay simple (just connectivity). Skills get smart (orchestration and domain logic).
Future Vision: Continuous Learning
Current agents don't learn across sessions. Every conversation starts fresh. Skills make learning transferable.
When you first start using Claude, this standardized format gives a very important guarantee. Anything that Claude writes down can be used efficiently by a future version of itself. This makes the learning actually transferable.
— Barry Zhang (00:13:35)
The Vision
"Our goal is that Claude on day 30 of working with you is going to be a lot better than Claude on day one."
Already Real
"Claude can already create skills for you today using our skill creator skill and we're going to continue pushing in that direction."
How It Works
- 1. Claude creates a skill to remember something important
- 2. Future versions of Claude can read that skill
- 3. Knowledge persists across sessions
- 4. The agent gets better over time
The Collective Knowledge Base
The vision that excites Anthropic most is a collective and evolving knowledge base of capabilities curated by people and agents inside organizations.
Network Effects
As you interact with an agent and give it feedback, all agents in your team and org get better too.
Instant Onboarding
When someone joins your team, Claude already knows what your team cares about.
Community Benefits
Skills built by others make your agents more capable, reliable, and useful.
"So just like when someone else across the world builds an MCP server that makes your agent more useful, a skill built by someone else in the community will help make your own agents more capable, reliable, and useful as well."
— Barry Zhang (00:13:00)
The Computing Analogy: Why Skills Are the Application Layer
In a rough analogy, models are like processors. Both require massive investment and contain immense potential, but are only so useful by themselves.
The OS made processors far more valuable by orchestrating the processes, resources, and data around the processor. A few companies build processors and operating systems, but millions of developers like us have built software that encoded domain expertise and our unique points of view.
— Barry Zhang (00:14:52)
Watch the computing analogy (00:14:41)| Layer | Computing | AI Agents |
|---|---|---|
| Hardware | Processors | Models (Claude) |
| OS | Operating System | Agent Runtime |
| Applications | Software | Skills |
The Opportunity
The real value in computing wasn't in building processors or operating systems—it was in the applications layer. Skills open up the "application layer" for AI development, where domain expertise gets encoded and shared.
Key Takeaways for AI Engineers
1. Intelligence ≠ Expertise
Agents are brilliant but lack domain knowledge. Skills bridge that gap.
2. Code is Universal Interface
You don't need domain-specific scaffolding. Code works for everything.
3. Progressive Disclosure Enables Scale
Load metadata first, full content on-demand. Thousands of skills possible.
4. MCP + Skills = Complete Stack
MCP provides connectivity, skills provide expertise.
5. Non-Technical Users Are Building Skills
Finance, legal, accounting teams are already doing this.
6. Enterprise Adoption is Real
Fortune 100s using skills for organizational knowledge and internal tools.
7. Continuous Learning is Possible
Agents can create skills for their future selves. Day 30 Claude > Day 1 Claude.
8. Skills Enable Collective Intelligence
One person's skill makes everyone's agent better.
9. Stop Building Agents, Start Building Skills
The future is composable, not monolithic. As Barry Zhang says: "So we think it's time to stop rebuilding agents and start building skills instead." (00:15:55)
Source Video
Don't Build Agents, Build Skills Instead
Barry Zhang & Mahesh Murag, Anthropic • AI Engineer 2024
Research Note: All quotes in this report are timestamped and link to exact moments in the video for validation. This analysis was conducted using multi-agent transcript analysis.
Technologies Mentioned: MCP (Model Context Protocol), Claude Code, Agent Skills, Python, Git, File Systems, Bash, Notion, Browserbase, Cadence