AI employee or agentic browser: how to tell them apart

What is an AI employee, precisely?

A chat-resident agent with a connector catalogue.

You mention it in a thread and it acts through connections you authorized: read the CRM, post the doc, update the ticket. The work runs in the vendor’s cloud, so it keeps going when your laptop shuts.

Connector catalogues in these products run from the hundreds into the thousands as of this writing, and that is maintenance nobody on your team ever has to do. Pricing is usually credits, sometimes per seat, sometimes per workspace.

What is an agentic browser, precisely?

An agent inside a desktop browser, working the sessions you are already signed into.

There is no connector step: the tab is already authenticated, so a companion sees what you see and clicks what you would click. That covers every site nobody built an integration for, which is most of the web.

The reach comes with supervision. In Strawberry, consequential actions pause on a card naming the action, you can watch the tab while it works, and Ask Always mode in Settings → AI permissions widens the gate to everything. Details at /security.

What is each shape built for?

Both lists describe terrain, and most companies own some of each.

• The AI employee is built for tools with decent APIs, decisions happening in Slack, a deliverable somebody reads, and jobs that have to finish overnight.
• Its cost shape is shared too: one workspace subscription rather than a licence per desk.
• The agentic browser is built for work behind a login, on a portal with no API, on the open web, or in a tool whose export button produces a PDF.
• It is also built to be watched: a tab you can interrupt mid-run, rather than a run history you read afterwards.

Sorting your own week

Count logins for a week. Write down every task you did by hand and mark whether finishing it needed a login no connector reaches. If it is most days, the browser side is where your unautomated hours are, and the catalogue of specific jobs at the agentic work index is a faster check than any feature grid.

Running both, which is what most people do

The chat agent owns the recurring API-to-API plumbing and the reporting the team reads. A companion owns the gated, judgement-heavy work that never got automated because nobody could automate it. Finished examples of the second kind are at /real-runs, and the architectural version of the split is at cloud browser agents vs. local agentic browsers.

Frequently asked questions