Working Title: Beyond Tool Calling: Agent Skills, MCP, and the New Programming Model for AI
—
In 2023, the big breakthrough was function calling — giving LLMs the ability to invoke external tools. In 2024, everyone built agents that could call APIs, search databases, and send emails. By 2025, the industry realized that raw tool calling wasn’t enough.
The problem: a function call is a single, stateless invocation. But real-world tasks require sequences of tool calls, context about when and how to use them, guardrails to prevent misuse, and verification that the task was completed correctly.
Enter agent skills.
The Evolution: From Functions to Skills
2023: Function Calling. „Here’s a list of functions. Call the right one with the right parameters.“ Simple, but limited. The LLM had no context about when to use each function or how to chain them.
2024: Tool Use. „Here are your tools. Figure out how to accomplish the task.“ More flexible, but the agent had to reason about tool selection from scratch every time. No reusable patterns.
2025-2026: Agent Skills. „Here’s a skill — a pre-packaged capability with tools, context, guardrails, and success criteria.“ The agent doesn’t just know what tools exist; it knows how to accomplish entire categories of tasks.
The February 2026 paper „SoK: Agent Skills — Beyond Tool Use in LLM Agents“ formalized this shift. The key insight: skills encode not just capability, but expertise.
What Is an Agent Skill?
An agent skill is a self-contained capability package that includes:
1. Trigger conditions — when should this skill activate?
2. Tools — what external capabilities does it need?
3. Context — what background knowledge is required?
4. Procedure — what steps should the agent follow?
5. Guardrails — what should the agent never do?
6. Verification — how does the agent know it succeeded?
7. Examples — what does good output look like?
Think of it as a „playbook“ for a specific type of task. Instead of the agent reasoning from scratch every time, it follows a proven pattern.
MCP: The Transport Layer for Skills
The Model Context Protocol (MCP) has emerged as the standard way to connect agents to external tools and data sources. In 2026, MCP is to agents what HTTP is to web applications — the universal protocol.
MCP matters for skills because it provides:
- Standardized tool discovery — agents can find and use new tools without code changes
- Portability — a skill built on MCP works across any MCP-compatible agent
- Composability — skills can call other skills through MCP
- Security — MCP defines authentication and authorization patterns for tool access
IBM’s 2026 Guide to AI Agents puts it bluntly: „Tool calling is table stakes. Skills, delivered through MCP, are the differentiator.“
The Skill Pattern Catalog
After analyzing hundreds of production agent deployments, four skill patterns emerge:
Retrieval Skills
Purpose: Find and synthesize information from external sources.
Tools: Search APIs, vector databases, document stores.
Guardrails: Always cite sources; never fabricate information.
Example: „Research the latest Gartner predictions for enterprise AI adoption.“
Reasoning Skills
Purpose: Break down complex problems and synthesize solutions.
Tools: Calculator, code interpreter, knowledge graphs.
Guardrails: Show your work; flag uncertainty; ask for clarification when ambiguous.
Example: „Analyze the cost-benefit of deploying 100 AI agents vs. hiring 10 engineers.“
Action Skills
Purpose: Execute tasks in external systems.
Tools: Email APIs, calendar systems, deployment pipelines, CRM.
Guardrails: Confirm before destructive actions; log all actions; require human approval for high-stakes operations.
Example: „Schedule a meeting with the engineering team and send calendar invites.“
Safety Skills
Purpose: Ensure the agent operates within acceptable boundaries.
Tools: Content classifiers, PII detectors, policy engines.
Guardrails: When in doubt, escalate to human; never bypass safety checks.
Example: „Before responding, check that the output contains no PII and complies with company policy.“
How to Design a Skill
The most effective skill format uses a structured markdown document (SKILL.md):
„`
—
name: research-and-summarize
description: Research a topic using web search and produce a structured summary
tools: [web_search, web_extract]
—
Trigger
Activate when the user asks for research, fact-checking, or „what’s the latest on X“
Procedure
1. Identify the core question
2. Search for recent, authoritative sources
3. Extract key information from top 3-5 sources
4. Synthesize findings into a structured summary
5. Cite all sources
Guardrails
- Never fabricate statistics or quotes
- Flag conflicting information from sources
- If no authoritative sources found, say so explicitly
Verification
- All claims are traceable to cited sources
- Summary is under 500 words
- At least 3 sources cited
„`
This format is human-readable, machine-parseable, and can be version-controlled. It’s the foundation of composable agent capabilities.
The Bottom Line
The shift from tool calling to agent skills is the most important architectural change in AI agent development since the introduction of function calling. Skills encode expertise, not just capability. They make agents more reliable, more maintainable, and more composable.
If you’re still building agents that reason about tool selection from scratch every time, you’re leaving performance and reliability on the table. It’s time to start building skills.
—
Word count: ~1,000 (excerpt — full draft would expand with more code examples, architecture diagrams, and case studies to reach 1,800-2,200 words)