The difference in one sentence
An agent receives a goal and decides how to get there. A skill receives a task and follows a procedure that is already written. The agent has its own role, context and criteria. The skill has instructions, and sometimes code, that work the same way every time they are called.
| Criterion | Agent | Skill |
|---|---|---|
| What it receives | A goal | A defined task |
| What it does | Decides the path | Follows a procedure |
| How much it varies | With every request | Almost not at all |
| How it is checked | Model judge or review | Test or fixed rule |
Where the rule came from
Experiment #007 compiled 14 research sources into a context engineering squad. The architecture ended up with 9 agents, 8 skills and 6 workflows. The most relevant finding was that splitting it this way is simpler and stronger than doing everything with agents. Each agent carries its own full frameworks, with no external references, and the mechanical work lives in the skills.
How to decide at each step
- Does the step require a choice? If two competent people would do it differently, it is agent work.
- Does the step repeat the same way? If the steps fit in a fixed list, it is a skill.
- How will you check it? If a test settles it, it is a skill. If it needs criteria and reading, it is an agent, with a separate check.
- When in doubt, fewer agents. That is what worked in #007: fewer agents, each more complete, and the repetitive work in skills.
The most common mistakes
- An agent for everything. Every mechanical step becomes an agent, and the system gains cost and handoff points where it can fail, as described in multi-agent systems.
- A skill that decides. The procedure piles up "if this, do that" until it becomes judgment. At that point, it is time for an agent.
- An agent without skills. The agent rewrites from scratch, every time, a procedure that could be written once.
Frequently asked questions
Is a skill the same as a prompt?
No. A prompt is a loose instruction. A skill is a packaged procedure, with instructions and sometimes scripts, that an agent calls when it needs it.
Can one agent use several skills?
Yes. It is the most common design: the agent decides what to do and calls the skills that execute each part.
Does this apply outside Claude Code?
Yes. The name changes from tool to tool, but the split between who decides and who executes a fixed procedure applies to any agent system.
Want to apply this to your operation?
The first call is a free one-hour diagnosis.