How do AI computer-use agents work, and how do you use one safely without giving it your passwords?

Computer-use agents run a simple loop — screenshot, think, click or type — and this week's news (Tencent's Hy4 with native computer use, Cohere's models on OpenRouter, Anthropic's server-side message threads) shows the tech going mainstream. The safe setup: a low-privilege account, approvals before payments and sends, step limits, and never your password vault.

You have probably seen the demos: an AI opens a browser, clicks around a store, fills in a form and checks out, with nobody touching the mouse. Short answer: a computer-use agent is a language model wrapped in a loop. It takes a screenshot of your screen, decides the next small action — a click at exact coordinates, a keystroke, a scroll — a helper program carries that action out, and then it screenshots again.

It repeats that loop until the task is done or it gets stuck.Why it matters: This stopped being a lab trick this week. On November 4, Tencent fully released Hy4, a 770-billion-parameter mixture-of-experts model with a one-million-token memory — Hunyuan's first model with built-in vision and native computer use across both regular apps and the command line. The same day, Cohere's Embed 5 and Rerank 5 models arrived on OpenRouter, giving agent builders search and retrieval tools through one API key.

And Anthropic's Claude API message threads, now in public beta, keep the conversation state on the server instead of resending the entire history every turn, cutting median time-to-first-token by about 35 percent on very long requests — which matters, because that resending is exactly what makes long agent tasks slow and pricey.How to use one safely, step by step:1. Give the agent its own low-privilege account — a separate browser profile or user account with no saved cards, no banking and no main email attached.2. Turn on approvals for anything irreversible — payments, sending messages, deleting files and changing settings should always pause and ask you first.3.

Set a step limit and a time limit, so a confused agent stops instead of clicking in circles.4. Never paste in your password vault, banking login or main email credentials. If a task needs a sign-in, you type it yourself.5.

Practice on a throwaway task first — book nothing, buy nothing — and watch how it recovers when a page looks different from what it expected.Common mistake: treating an agent like a chatbot. A chatbot's worst day is a wrong sentence; an agent's worst day is a wrong action. Add prompt injection — web pages can hide text that tries to talk your agent into doing something you never asked for — and approvals stop being optional.Ad’s take: Start agents on research and busywork inside a sandbox account.

The tech is real now, but your passwords stay yours.