Agents vs Skills

Core Difference

	Agent	Skill (Plugin)
Role	Decides what to do and when	Does what it’s told
Reasoning	Yes — plans, reacts, adapts	No — deterministic execution
Initiates actions	Yes	No
Calls other things	Yes (calls skills/tools)	Usually no
State/memory	Maintains context across steps	Stateless by default
Complexity	High — open-ended goals	Low — fixed inputs/outputs

Mental Models

The Contractor & Power Tools

An agent is a contractor who decides how to complete a job — which steps to take, in what order, and how to react when things go wrong. Skills are the power tools the contractor picks up and uses. The tools don’t decide anything; the contractor does.

The Manager & Employees

An agent is a manager who receives a vague goal (“increase sales”) and figures out the plan. Skills are specialist employees who each do one thing well when asked — the accountant runs the numbers, the designer makes the mockup. The manager coordinates; the employees execute.

The Function Test

If you can write it as a deterministic function, write it as a skill. Only reach for an agent when the decision-making itself is the hard part.

When to Create a Skill

• The task has a predictable input/output — same input always produces same output
• No reasoning required — just execution (fetch a URL, query a DB, send an email, parse a file)
• It’s a reusable capability that multiple agents might need
• You can unit test it in isolation without an LLM
• The task is a leaf node — it doesn’t need to call other skills to complete

Examples:

• GetWeather(city) → returns current weather
• SearchDatabase(query) → returns matching records
• SendEmail(to, subject, body) → sends an email
• ParseInvoice(file) → extracts line items from a PDF

When to Create an Agent

• The task requires deciding which steps to take and in what order
• The goal is open-ended — the path to completion isn’t known upfront
• It needs to react to intermediate results before deciding the next action
• Requires multi-turn reasoning, memory, or conversation state
• The task involves error recovery — retry, reroute, or ask for clarification when something fails
• Multiple skills need to be coordinated to achieve the goal

Examples:

• ResearchAssistant — decides whether to search the web, summarise findings, ask a clarifying question, or cite sources based on what it finds
• DatabaseAssistant — determines which tables to query, validates the generated SQL, retries on failure
• InvoiceProcessor — parses the invoice, looks up the vendor, checks for duplicates, flags anomalies, routes for approval

Decision Framework

Can you write it as a function?
├── Yes → Is it reusable across multiple agents?
│          ├── Yes → Skill
│          └── No  → Inline logic (don't abstract yet)
└── No  → Does it need to coordinate multiple steps?
           ├── Yes → Agent (possibly with sub-skills)
           └── No  → Revisit — you may be over-engineering

Scenario Guide

Scenario	Use
Fetch data from an API	Skill
Generate a SQL query from natural language	Skill (wraps LLM with fixed I/O)
Answer a question that may require multiple lookups	Agent
Send a notification	Skill
Triage an inbox and decide what to do with each email	Agent
Convert a file format	Skill
Autonomously complete a multi-step research task	Agent
Validate a form field	Skill
Manage a workflow with branching logic and retries	Agent (or Workflow)

Common Mistakes

• Making agents too granular — if your “agent” only ever calls one skill, it should just be a skill
• Making skills too smart — if your skill is making decisions about what to do next, it’s becoming an agent
• Skipping skills entirely — agents that inline all their logic are hard to test and reuse
• Agent for everything — most tasks are deterministic; adding an LLM where a function suffices adds latency, cost, and unpredictability

In Specific Frameworks

Framework	Term for Skill	Term for Agent
Microsoft Agent Framework	Tool / MCP Server	Agent
Semantic Kernel	Plugin (formerly Skill)	Agent
LangChain	Tool	Agent / Chain
Claude Code	Skill (slash command)	Claude (the model + harness)
AutoGen	Tool / Function	ConversableAgent

See Also:: Microsoft Agent Framework
See Also:: Agentic Frameworks
See Also:: Semantic Kernel

---

📇 Additional Metadata

• 🗂 Type:: note
• 🏷️ Tags:: ai
• 📡 Status:: #status/🌱