AI - Agent Skills
In the context of AI agents, a Skill is a discrete, reusable capability that an agent can perform to accomplish a specific task.
While the industry has used terms like "Tools," "Plugins," "Actions," and "Extensions," the word Skill is becoming the preferred "human-centric" term. If an agent is a "Digital Employee," then a "Skill" is a line item on its resume (e.g., "Can process invoices," "Can search GitHub," "Can book flights").
What exactly is a "Skill"?
A skill is essentially a "wrapper" around a piece of code or an API. It consists of three parts:
- The Name: (e.g.,
search_knowledge_base) - The Description: This is the most important part. It tells the LLM why and when to use this skill. (e.g., "Use this skill when the user asks a question about company health insurance policies.")
- The Logic: The actual code that runs, hits an API, or queries a database.
Is "Skill" replacing concepts like MCP?
No. They are two different layers of the same stack.
Think of it like this:
- MCP (Model Context Protocol): This is the Connector (the USB-C port). It is the technical standard that allows the agent to talk to a data source.
- Skill: This is the Capability (the software). It is what the agent does once it is connected.
The Relationship: You use MCP to connect your agent to your SQL Database. Once connected, your agent gains the Skill of "Database Querying."
MCP makes skills portable. In the past, if you wrote a "Salesforce Skill" for an OpenAI agent, it wouldn't work for a Claude agent. With MCP, you write the Salesforce connection once, and it becomes a "Skill" that any MCP-compatible agent can use.
"Skill" vs. "Tool" vs. "Extension"
The terminology is messy because different companies use different words for the same thing:
| Term | Who uses it? | Connotation |
|---|---|---|
| Tool | OpenAI / Anthropic | Very technical. Usually refers to a single function or API call. |
| Extension | Google (Gemini) | Refers to connecting the AI to another app (like Gmail or Maps). |
| Skill | Microsoft / Amazon / CrewAI | Higher level. Suggests a "competency" (e.g., "The skill of market research"). |
| Action | OpenAI (GPTs) | Specifically refers to an agent performing a task in a third-party app via API. |
Why the shift toward the word "Skill"?
The industry is moving toward "Skill" for two reasons:
- Composition: A "Skill" is often a collection of "Tools." For example, a "Travel Planning Skill" might use a "Flight Search Tool," a "Hotel Booking Tool," and a "Weather Tool" all at once.
- Agentic Identity: As agents become more autonomous, we describe them in human terms. We don't say a human employee has a "Browser Tool"; we say they have "Research Skills."
How you "build" a Skill
If you are using a managed platform like Google ADK or Claude Managed Agents, building a skill usually looks like this:
- Define the Schema: You write a JSON object describing what the skill does and what inputs it needs (e.g., it needs a "Start Date" and "End Date").
- Connect the Implementation: You provide the URL of the API or the Python code that should run.
- Instruction: You give the agent a "Playbook" instruction on how to use that skill effectively (e.g., "Always verify the price before using the 'Book_Flight' skill").
Summary
- MCP is the "how" (the protocol).
- Skill is the "what" (the capability).
- Tool is the "unit" (the specific function).