Opening the Box
In the last level we laid out a spectrum: chatbot at one end, assistant in the middle, agent at the far end. So what actually moves an AI along that line?
It comes down to four ingredients: goals, tools, memory and autonomy. Think of them as four dials. Turn any one of them up and the AI acts a bit more like an agent. Turn them all up and you have something that can take on a whole job for you.
Let's take them one at a time.
Ingredient 1: Goals
A plain chatbot answers the question you just asked, and that's it. Each reply is a single step.
A goal changes that. Instead of "What's a good hotel in Lisbon?" you hand over an outcome: "Find me three hotels in Lisbon under 150 euros a night, near a tram stop, and compare them." Now the AI has to work out the steps for itself. It searches, filters, compares and then knows when it's done.
Here's an easy test. Does the AI have a finish line it's working towards, or is it just replying to your last message? A finish line means the goal dial is turned up.
Ingredient 2: Tools
On its own, a language model can only produce text. Tools let it reach outside the chat box.
A tool might be web search, a calculator, reading a file you've uploaded, opening a web page or sending an email. Picture a very bright new colleague. Without tools, they can only talk to you. Give them a laptop, a phone and access to your calendar, and they can actually do things.
That's the big jump. Tools are usually where an AI gains the most new ability, because they turn "here's what you could do" into "I've done it." We'll dig into what changes when an AI acts in the next level, so for now just remember that tools are the hands.
Ingredient 3: Memory
Memory is what the AI carries from one moment to the next.
There's short-term memory, meaning what it holds onto during one conversation or task. There's also longer-term memory that lasts across sessions. ChatGPT, for example, has "saved memories", which are things you've asked it to remember, like "I'm vegetarian when you suggest recipes." It can also draw on your past chats. Gemini keeps similar info in what it calls "Personal context". In both, you can view, edit or turn these off in settings.
Why does memory matter for agents? A task with ten steps falls apart if the AI forgets step two by step seven. And memory across sessions means you don't have to explain your preferences every time.
There's a catch, though. Memory can also hold onto something wrong or out of date, and that mistake then gets carried into everything that comes after.
Ingredient 4: Autonomy
Autonomy is how much the AI does without checking with you first.
It helps to picture a ladder:
- It only answers. You do everything.
- It suggests an action and waits for your OK.
- It acts, then reports back.
- It acts on its own, even on a schedule while you're away.
You'll see all of these rungs in real products. ChatGPT, for instance, can run scheduled tasks, either one-off or recurring, so it can do something at a set time without you being there. Its agent mode is designed to ask your permission before consequential steps like sending emails or making purchases. That's a deliberate choice to keep autonomy lower exactly where the stakes are higher.
Let's Rate a Task Together
Let's try it. Say you ask an AI: "Every Monday morning, check three news sites for updates about my industry and email me a five-point summary."
How would you rate each dial?
- Goals: high. It has a clear outcome and a finish line.
- Tools: medium to high. It needs web browsing and the ability to send email.
- Memory: medium. It needs to remember your industry and which sites to check, and ideally what it already told you last week.
- Autonomy: high. It runs every week with nobody watching.
Now compare that with "Summarise this article I pasted." The goal is small, it uses no tools, it needs no memory and it barely has any autonomy. That's assistant territory.
Capability Versus Risk
Here's the key insight. The ingredients that add the most capability are often the ones that add the most risk.
Tools give the AI hands, and hands can do damage as well as good. Autonomy takes away the moment where you'd notice a problem. Put them together and you get real power, plus real room for mistakes.
So when you look at any example, ask yourself two questions. Which dial gives this AI its new ability? And which dial would hurt most if something went wrong?
Often the answer is the same: a powerful tool, used without asking. Letting an AI search the web is fairly low-stakes. Letting it pay for things without checking with you is a different story.
Try This
Pick something you already use AI for. Rate it low, medium or high on each of the four dials. Then ask: if I turned one dial all the way up, which one would I want to keep an eye on?
Key Takeaways
- Goals give the AI an outcome to work towards, not just a question to answer.
- Tools let it act beyond text, like searching, browsing or sending. This is often the biggest jump in ability.
- Memory carries information across steps and sessions. That helps, but it can also carry mistakes forward.
- Autonomy is how much it does without asking you first, from "only answers" to "runs on a schedule".
- The most risk usually sits where a powerful tool meets high autonomy.