Every huge companies are building personal agents right now. The part I think matters most is the two-agent structure, so that's what Nico is built around.

The front agent talks to you. Always available, only one session, one memory, continuously compaction, and it carries almost no tools, so it can answer fast.

The worker agent does the job. It picks up tasks the front agent puts in an inbox and reports back when they're done. It almost never talks to you; anything it needs goes through the front agent.

I think this is the real step forward, but the attention goes to the cloud computer. I get why: you can hand the agent a task from your phone. But, it also makes everything more complicated. People need to upload the entire environment while

So I built one with the cloud computer taken out. Nico is a native Mac app. You hold the right ⌘ key (or fn) and talk; the front agent answers by voice; the worker agent is pi (github.com/earendil-works/pi) runs by your own AI subscription, in a sandbox.

With computer-use (via cua-driver), browser-use (via a browser extension), and a built-in browser, she can clean up your desktop, or answer "why is my Mac so hot". If you say "this", she looks at the screenshot of the active window. Most tasks run without you; sending an email waits for your click.

The downside: when you're not at your Mac, you can't use it. I haven't figured out what to do about that.

Would you want a personal agent that only lives on your Mac?