Agent, not chatbot

A chatbot answers and stops. A personal AI agent keeps going: you give it a goal, and it takes actions in the world to reach it — opening a browser, filling forms, sending messages, making purchases — often continuing after you've closed the app.

Strip away the branding and all three of 2026's big personal agents are built from the same four parts:

1. A model that reasons and decides (Grok Bot runs on grok-4.7; Meta Muse on Muse Spark, with newer Muse Wonder models; OpenAI Dots on GPT-6 Astra).
2. Its own computer in the cloud, usually with a browser, where it actually does the work.
3. Connections to your apps — email, calendar, payments, and so on — that you grant.
4. A goal and a loop: it works toward what you asked, checking back when something changes or it needs approval.

The model is the part you'd assume matters most. It isn't. Every one of these models is capable enough to do real damage and real good; what decides whether an agent is a help or a liability is everything around the model — the computer it runs on, the apps you connect, the limits on those apps, and the approvals in front of risky actions.

That's the whole mindset of this course: you are not tuning a model, you are running a worker. A worker with your logins, your card, and the ability to act while you sleep. Treat it like one — give it exactly the access the job needs, make it check before anything irreversible, and watch what it costs — and a personal agent is genuinely useful. Skip that, and it's a quiet risk.

The good news: the habits that make Grok Bot safe and cheap are the same ones that make Muse and Dots safe and cheap. Learn them once here and they transfer to whatever ships next.