The Three-Layer Model of Agents
Swipe to show menu
Every action-taking agent, no matter how complex it looks, is made of three parts. Once you see the parts, the whole space stops feeling magical and starts feeling buildable.
This is the LLM itself: Claude, GPT, Gemini. The brain decides what to do. It reads your request, plans a sequence of steps, and figures out which tool to reach for at each step. Without a brain, nothing thinks. But a brain alone can only talk.
Tools are the things the brain can actually use. Sending an email is a tool. Searching the web is a tool. Adding a row to a spreadsheet is a tool. Posting to Slack is a tool. When you "connect" an app to your agent, you're handing the brain a new pair of hands. More tools means more things the agent can do.
Some tasks don't have clean tools. There's no "book a flight" button you can hand to the AI — flights live behind a website. So agents need a way to see and click, the same way you do. This is where browser agents like Comet and Claude for Chrome come in. They look at the page, read what's on it, and click around. Eyes are slower than hands, but they work on anything.
A real workflow uses all three.
Imagine this request: “Find the three most-mentioned competitors in my last 50 emails and add them to my CRM.”
The brain plans the steps. The hands read the emails through a Gmail connector. The brain analyzes the text. The eyes open the CRM website, because there is no clean tool for it, and click through the “add company” form.
You do not need to build any of this yourself. You just need to understand which layer is doing what when something works, and which layer is broken when something does not.
Thanks for your feedback!
Ask AI
Ask AI
Ask anything or try one of the suggested questions to begin our chat
