Agent browsers versus operator profiles

The words are collapsing. Vendors say "agent browser" and "antidetect profile" in the same sentence, as if a planner and a folder were one product. They are not. This comparison exists so the names stop sharing a coat.

The commercial warning is AI agents in a browser they do not own. This page is the architecture.

Operator profile

An operator profile is a local browser identity. On this desk that means a Chromium or Firefox user-data directory, a cookie store, optional proxy credentials, and a fingerprint seed that stays put across launches. A person starts the profile. A person types. The disk is the operator's Windows box.

MaskWright 0.1 is a launcher for that object. No account. No cloud copy of the cookies. No planner.

Agent browser

An agent browser is a runtime a model can drive. The model receives a goal. It chooses clicks. The browser may be local or hosted. The control plane is not the person. The session is instrumented for the planner, which means logs, screenshots, and often a vendor-side copy.

Hosted variants are cousins of Browserless and Browserbase are not operator profiles. A protocol that lets a model call tools is not a folder either. MCP is not a local profile.

The comparison that matters

The parent essay is browser automation versus a real profile. An agent browser sits on the automation side even when the UI looks friendly.

What I will not merge

I will not add a prompt box that drives profiles. I will not let an unofficial social helper in through the agent door. Social media automation tools we refuse. LinkedIn automation tools we refuse.

I will not let the agent solve a captcha and call it research. Scripts, including planners, stay outside work profiles. Local scripts stay outside work rooms.

What each architecture is good at

Operator profiles are good at isolation. Several authorized profiles on one PC. Separate cookie stores. A person who can explain every click. Backups you control. A day that still works if the vendor site is down, as long as you do not need ads, updates, or geo checks.

Agent browsers are good at throwaway tasks on hosts that invited a machine. They are fast to demo. They are bad at custody. They are bad at 2FA that should stay on a phone you hold. They are bad at any login whose loss is a business event.

If a vendor sells both in one download, ask which object you are buying. If the answer is "both," you are buying the blur.

How to keep the names honest

If you are sitting in a window, closing it, and coming back tomorrow to the same cookies, you are in an operator profile.

If you are typing a goal and watching a pane click, you are in an agent browser. Treat that pane like a contractor with a screen share. Do not hand it the keys to a store.

The mix-up happens because both objects show Chromium. The chrome is not the architecture. Ask who holds the folder, who plans the next click, and who can replay the session after you close the lid. Those three answers decide the name.

The ethics line is automation ethics we follow. The cluster is Automation. Use this page when a landing page tries to sell you both jobs as one download. Compare local versus cloud if custody is the buying question.

FAQ

Can an agent use MaskWright profiles?

Not in 0.1. There is no prompt box and no tool listing for profiles. Paste text into a model if you want a summary. Do not hand it the folder.

Is a local Chromium that a model drives still an operator profile?

No. If the model picks the clicks, it is an agent browser even when the window is on your desk.