What an agentic browser is
An ordinary browser displays pages and leaves the decisions to the person. An agentic browser contains an AI agent you can hand a task to in plain words: “find these running shoes cheaper and arrange delivery by Friday”. The agent opens sites, follows links, picks a size, fills in forms and either stops before payment or carries on — depending on the settings. In essence it is computer use built into the user’s own browser.
Examples
| Browser | Company | Launched | What the agent does |
|---|---|---|---|
| Comet | Perplexity | 9 July 2025 (Max plan), free for everyone from 2 October 2025 | Completes tasks on websites, including purchases |
| Opera Neon | Opera | 30 September 2025; open subscription at $19.99 a month from 11 December 2025 | Neon Do opens tabs, compares sites, fills in forms |
| Copilot Mode in Edge | Microsoft | 28 July 2025, experimental mode | Compares products across open tabs; agentic Actions in preview from October 2025 |
| Yandex Browser with Alice AI | Yandex, the dominant search engine in Russia | December 2025, early access | The “Find cheaper” agent compares prices for the user’s region; one-click purchase via Yandex Pay |
OpenAI’s ChatGPT Atlas launched on macOS on 21 October 2025 and was shut down on 9 August 2026, with its agentic features moving into the ChatGPT app and a Chrome extension.
What it means for a store
- The visit looks human. The agent runs in a real browser, often inside the user’s session, so it lands in AI agent traffic that is hard to separate out: it skews conversion, behavioural metrics and A/B tests.
- CAPTCHAs and bot protection. The agent stumbles on them, and the order is lost together with the shopper who delegated it.
- Someone else’s instructions. Hidden text on a page, a review or a product description may carry a prompt injection that the agent will execute.
Important: in August 2026 the US Court of Appeals for the Ninth Circuit, in Amazon v. Perplexity, found a violation of the Computer Fraud and Abuse Act unlikely: when a user gives an agent a task, it is the user who accesses the site, with the agent only helping. The legal picture is still forming, but blocking every agent outright also means losing real customers.
How to prepare
- Walk through your own checkout in two or three agentic browsers and note where the agent gets stuck — a direct test of your agent experience.
- Keep price, stock, sizes and delivery terms in text and structured markup, not only in images and pop-ups.
- Put sessions that look automated into a separate segment and, where possible, exclude them from A/B tests.
- Moderate reviews and marketplace-seller descriptions for hidden instructions.
- Give agents a stable path instead of clicking through the layout — an API or Web MCP.