Subscribe

What's Inside an AI Agent? The Five Parts Every Agent Needs

What is an AI agent?

An AI agent is an AI model with a job, tools, skills, memory and a schedule. When I use ChatGPT in a chat window, it waits for me. An agent keeps working after I close the window. It can check my calendar, read a doc or run a task on a schedule, then tell me what it did.

The model underneath is the same kind of AI you already chat with. What makes it an agent is everything around it.

What are the five parts of an AI agent?

Job and personality

These are written instructions for who the agent is, what it does and how it talks. The tighter the job, the better it works. My agent Byline drafts my newsletter and nothing else. My agent Eddy is my front door: I talk to him, and the other agents report to him.

Tools and access

These are the accounts and apps it can use, like your calendar, inbox, a browser or Notion. You decide what it can reach. My finance agent handles Stripe, Xero and invoices. My agents can sign in with logins from 1Password, and when one needs a secret, I type it into a masked field so the agent never sees the value.

Skills

A skill is a saved how-to it follows the same way every time. Byline has skills for my newsletter structure and for editing out AI-sounding phrases. In Grok Bot, I can also record a task in the browser once, and it can do it again.

Memory

Memory lets the agent remember what you told it last week, so you don't repeat yourself. When I ran my main agent on Hermes, the part I loved most was that it learned from its own work.

Schedule

A schedule or trigger tells the agent when to start without you. Every morning, a routine scans the next two weeks on my calendar and adds travel blocks with Google Maps links. If nothing changed, I don't hear about it.

How is an agent different from ChatGPT?

From the outside they look alike. You talk to both in plain English, and both get things wrong sometimes, so you still check the work.

The difference is who keeps the work moving. In a regular ChatGPT chat, you ask, read the answer and ask again. When you close the window, the work stops. An agent has a job, its own tools and the access you give it. You tell it what you want and walk away.

The test is simple: close the window. If the work keeps happening, it's an agent.

How the products map to the five parts

Most agent products now use these same parts, with their own names. Here's what each one's official page shows.

  • Notion Custom Agents have instructions, a "Tools and access" section and triggers, including recurring schedules. They can also use skills.
  • Hermes Bot Mode gives each bot its own role, model, memory and skills, plus recurring routines.
  • Grok Bot bots get a name, title and description, connect through MCP servers, plugins and skills, and run on routines and triggers.
  • ChatGPT Work connects to your apps through plugins and runs Scheduled Tasks.
  • Claude Cowork uses plugins that bundle skills, connectors and sub-agents, and you can schedule tasks.
  • dots learn from feedback over time and connect to apps through plugins.
  • Muse remembers what matters to you, and you can tell it to forget things.

For a side-by-side of these products, see ChatGPT Work vs. Claude Cowork vs. dots.

Where agents run: in the cloud, on a schedule

Many agents keep going with your laptop closed because they run in the cloud. Grok Bot, dots and Muse each run on their own computer in the cloud. Claude Cowork keeps going after you close your laptop, and ChatGPT Work's Scheduled Tasks keep projects moving when you're away.

The schedule is the other half. A routine runs at a set time, like my morning calendar check. A trigger runs when something happens, like a new email or a new row in a database.

What I'd give an agent first

  1. Pick one annoying, repeated job, and give it to one agent.
  2. Write its job in a few plain sentences, including what it should never do.
  3. Start with drafts or a test you can check. You approve before anything goes out.
  4. Add a schedule once you trust it, then add the next agent.

My rule is that agents draft and I send. Nothing goes out without my yes.

If you want help setting this up for your team, that's part of my fractional CTO work.

FAQ

Is ChatGPT an AI agent?

Regular ChatGPT chat isn't. It waits for your next message. ChatGPT now includes agents, though: ChatGPT Work can stay with a project for hours, and dots are always-on agents inside ChatGPT.

Do I need to code to build one?

No. Most agent products let you describe the job in plain English. Notion can draft a Custom Agent's instructions and triggers from a chat, and a Grok Bot is set up with a name, title and description.

What is an agent skill?

A saved set of instructions for one kind of work. Notion describes skills as "reusable instructions that tell agents how to do a specific kind of work." You write it once, and the agent follows it whenever that work comes up.

What is an agent harness?

The software wrapped around the model. It gives the model its tools, runs the steps and keeps track of the work. Grok Bot, for example, hands coding work to a specialized coding harness inside Cursor.

Is it safe to give an agent my accounts?

It can be, if you set limits. Give it only the access it needs, and require approval for anything that sends, buys or deletes. Most products build this in. Cowork asks before acting when you turn that on. Muse checks with you before an email or purchase. Grok Bot uses a separate review agent with allow and block lists.

Get the next one in your inbox. Subscribe to Automate With Rob. I define the terms in my Operator's Glossary.