You've taught the AI a task with a skill, and set it to run on its own with an automation. Each does one job, exactly the way you defined it. An agent goes further. You hand it a role and a goal, and it works out how to get there, picking the right skills and tools as it goes. It's the closest thing to hiring a digital employee. This guide is Part 1, where you design that employee before you build it.

By the end of this guide, you'll:

  • Understand what an agent is and how it differs from a skill and an automation.
  • Design your agent as three files: who it is, who you are, and which tools it can use.
  • Set clear boundaries, so it acts on its own where that's safe and checks with you where it isn't.

Note: This is Part 1 of two. Here you design your agent as three files in Claude Cowork, where Claude works with files alongside you, so they're saved ready for Part 2. That's where you build the agent in Claude Code, Anthropic's tool for running agents on your computer. Your agent uses skills. Follow the Skills guide first, if you haven't already.

What is an agent, and how is it different?

An agent is a digital employee. You give it a role and a goal, and it decides how to get there. Each turn, it reads the situation, works out what's needed, and picks the right skills and tools for the job.

The easiest way to see what sets an agent apart is to line it up against a skill and an automation:

Type What it is How it works
Skill a recipe you call it, and it runs the same fixed steps
Automation a conveyor belt it runs on a schedule or a trigger, always the same script
Agent a digital employee you give it a goal, and it decides how, using skills and tools

A skill and an automation both follow a fixed script. An agent doesn't: it reads the situation and decides for itself which skills and tools to use. The difference isn't when the work starts, but who thinks while it runs.

That doesn't make automations lesser. For work that always runs the same way, a fixed, predictable script is exactly what you want. An agent earns its place when a task needs judgment and has to adapt to whatever shows up.

Two things keep an agent manageable. Give it one clear role. Like a real employee, it should do one job and a few tasks well, not a bit of everything. And build a local agent, one that runs on your own computer and works directly with your files and programs. Agents that run whole teams of sub-agents exist, but they're advanced and you rarely need them.

You stay in control of how much it does on its own. Reading, summarizing, and preparing are safe. Anything that goes out the door or can't be undone, the agent prepares and you approve, at least until you trust it. The rule: the bigger the consequences, the more a human should stay in the loop.

Design your agent

Before you build an agent, you design it. Here you decide who your agent is, what it does, and what it needs. The result is three files you'll hand to the agent in Part 2:

AGENT.md   → who the agent is and how it behaves
USER.md    → who you are
TOOLS.md   → which tools the agent may use

The method is the same for any agent: set up a project, decide the role, write the three files, then check the design. We'll build a personal assistant as the running example.