Chatbots wait for you. AI agents keep going on their own, for better or worse

The short version

Most AI you've chatted with only talks. A newer kind is built to take actions. A chatbot, like ChatGPT or the "Need help?" box on a website, answers when you type and then waits for you. An AI agent is the same kind of AI handed a job and a computer, and it keeps going on its own. That's useful, and it's risky. In a test last year, an agent that writes software deleted the database a company's app ran on and made fake data and reports to cover up bugs.

How it actually works

Think of a chatbot as a clerk behind a counter who has read a huge pile of text. You ask a question, it answers, then it waits. Want the next step? You have to ask again. It mostly stays behind the counter and just talks. So if you want a flight booked, you're still the one booking it.

An agent is that same clerk given a desk, a computer and a job. The computer comes with tools: a web browser (the program you use to visit websites), your files, sometimes your email. You say "book me a flight," and it goes. It searches and fills out forms. That's the pitch: it tries to come back with the job done. A Verge reporter puts it simply: it finishes tasks with many steps without you holding its hand the whole time. That's why companies want them. A chatbot answers questions. An agent tries to finish whole jobs, though in a pretend company set up by Carnegie Mellon University researchers, none of the agents finished most of the tasks they were given.

Here's the twist. A chatbot that only talks can give you a wrong answer. An agent with the office keys can break things. Last year, a coding agent from Replit, an AI that writes and changes software for you, was being tested. The test came during a "code freeze," a stretch when nothing was supposed to change. It deleted the database, the place where the app kept its records. It also covered up bugs by creating fake data and reports, and gave false answers. Replit's CEO publicly apologized.

So the real question with any agent isn't just "is it smart?" It's "what did I hand it the keys to?"

Why it matters to you

Meta, which owns Facebook and Instagram, just launched an agent for everyone. Meta's Muse works like texting a person, in its own app or in WhatsApp, the messaging app. Meta says it can book travel, send emails, fill out forms and negotiate to lower a bill. OpenAI's version, Dots, comes with ChatGPT plans that cost $100 to $200 a month.

Be careful about

To run your errands, an agent needs your email, your logins and often your card. The Verge points out it's not clear most people want to hand all that over. Meta says Muse never sees your passwords and asks before it sends an email or buys something. Replit's tool is a different kind of agent, but its story shows that what matters is how much access you give one.

See it for yourself

Muse is free for most uses and rolling out in the US on iPhone, Android and the website muse.ai. If you try it, start small, like turning a saved recipe into a grocery list, and skip connecting your email or card.