Learn · Automate · 6 min
Agents that act for you: what to hand over, and what not to
ChatGPT agent, Claude in Chrome, Gemini and Meta's Muse can now click, type, book and buy for you. Here is how to get the time back without handing over the keys.
What makes it an agent
A chat answers you. An agent does things: it opens websites, fills in forms, sends emails and can pay. Anthropic's engineers define it as a model that decides its own steps and which tools to use, rather than following a fixed recipe. That freedom is what saves you time, and what makes it risky.
The ones you can use today
ChatGPT agent (since July 2025) works on its own virtual computer and asks before anything important. Claude in Chrome works inside your browser on calendars, replies and expenses. Gemini can use a remote browser, add to a shopping cart and pay with Google Pay. Meta's Muse (September 2026) sends emails, books travel and checks out with single-use cards. OpenAI's "dots" (late September 2026) run all the time in their own cloud computer.
The risk no one has solved: prompt injection
An agent reads web pages, emails and documents, and it cannot reliably tell your instructions from words hidden in what it reads. The UK's National Cyber Security Centre says this may be worse than older kinds of hacking. Researchers have hijacked agents in Perplexity's Comet browser, in several AI browser assistants at once, and in Muse within weeks of launch.
Start with errands that cannot hurt you
Good first jobs: research and compare, draft but do not send, fill a form you review before submitting, tidy a shopping list. Keep payments, banking, health records and anything you cannot undo for yourself, or make the agent stop and ask first.
Set it up safely, then check it saved time
Give it its own browser profile without your saved passwords. Connect only the accounts the task needs, and disconnect them after. Turn on the safety modes on offer, such as ChatGPT's Lockdown Mode, which switches off browsing and downloads when you do not need them. Read what it did before you confirm. Time how long the job took you last week, then time the agent plus your checking. If checking takes longer than doing, take that job back.
Try it now
Pick one weekly job and say it in a line: Ontold can help you plan which parts to hand to an agent.
Reviewed from: Building effective agents (Anthropic), Introducing ChatGPT agent (OpenAI), Piloting Claude in Chrome (Anthropic), Prompt injection is not SQL injection (UK National Cyber Security Centre), Indirect prompt injection in Perplexity Comet (Brave), BragJack attacks hijack AI browser agents (BleepingComputer), Muse suffers zero-day (Tech Times), ChatGPT Lockdown Mode (The Next Web), Gemini Apps Privacy Hub (Google), Muse, Meta's new personal AI agent (WIRED). Written in our own words.
Drafted with AI from the linked sources. Check the links for the full detail.