Agentic Browsers and Shopping Agents: What Happens When AI Clicks Instead of You

Every major AI lab shipped an agentic browser or a task-automation agent within about a year of each other, and most of them have already been shut down, reworked, or folded into something else. That churn isn’t a sign the idea failed — it’s a sign the idea is genuinely being figured out in public, fast, with real consequences for anyone building a site assuming humans are the ones clicking through it.

The timeline, company by company

OpenAI launched Operator as a research preview on January 23, 2025, initially restricted to $200/month ChatGPT Pro subscribers — an AI agent designed to perform multi-step tasks autonomously inside a browser environment. It was deprecated and folded into a broader “ChatGPT Agent” capability by August 31, 2025, a lifespan of a little over seven months as a standalone product.

OpenAI followed with ChatGPT Atlas, a full browser (not just an in-browser agent), launched October 21, 2025 — macOS-only at launch, built on Chromium, with an “Agent Mode” gated to Plus, Pro, and Business subscribers. Reporting indicates OpenAI began sunsetting Atlas by August 9, 2026; Windows, iOS, and Android versions were never shipped at all.

Perplexity’s Comet browser had a more gradual rollout: a limited release to Perplexity Max subscribers on July 9, 2025, followed by a free worldwide rollout on October 2, 2025 — Perplexity’s version has, as far as I could confirm, had more staying power than OpenAI’s Atlas.

Google’s Project Mariner started as a research prototype on December 11, 2024, then got a broader rollout announcement at I/O on May 21, 2025. Google reportedly shut it down on May 4, 2026, folding its capabilities into Gemini Agent and AI Mode rather than continuing it as a separate named product.

Four major agentic browser or task-agent products, launched within about ten months of each other, and by mid-2026 at least two had already been discontinued as standalone products. This space is moving faster than most sites’ technical infrastructure is prepared for.

I’ll flag directly: the 2026 shutdown dates for Atlas and Project Mariner come from tech-news reporting close to or after the point my own research could fully verify against a primary company announcement — treat those specific dates as reported, and worth double-checking against OpenAI’s or Google’s own current documentation if precision matters for your purposes.

What “an AI clicks instead of you” actually means practically

The shared idea across all four products is an AI agent that navigates a real browser, reads real pages, and takes real actions — filling forms, comparing options, completing purchases — on a person’s behalf, rather than just returning a text answer. That’s a fundamentally different kind of visitor than either a human reader or a traditional search crawler:

• It needs to parse your page’s actual functional elements — buttons, forms, navigation — not just extract readable content the way a search-indexing crawler does. A site that’s visually clear to a human but structurally messy in its markup may be harder for an agent to successfully act on, even if it renders fine.
• It’s making decisions on someone else’s behalf, often without them watching each step. That raises the practical importance of your page’s information actually being accurate and current at the moment of the interaction — a human might notice a stale “in stock” label and double check; an agent may simply act on it.
• It’s a new category of bot traffic that doesn’t map cleanly onto either “search crawler” or “human visitor” in your analytics, which makes it easy to under-measure or misattribute — similar to the broader AI-referral tracking problem across the AI-search space generally.

What’s actually worth doing about this today

Given how quickly this category is still churning — real launches, real discontinuations, all within about eighteen months — I’d resist over-investing in optimizing specifically for any one named agent product right now. What’s more durable:

1. Keep your site’s structural markup clean and semantic — proper form labels, clear button text, logical heading structure. This helps any current or future agentic browser parse and act on your page correctly, regardless of which company’s product is doing the parsing.
2. Keep critical information (price, availability, terms) accurate at the page level, not just in a separate feed — an agent acting directly on page content needs that content itself to be trustworthy, not just a data feed elsewhere that a human might cross-reference.
3. Watch your server logs for unfamiliar user-agents, even ones that don’t yet show up in your analytics platform’s default categorization — this is a genuinely new and fast-moving category, and default analytics tooling is likely still catching up to it.
4. Don’t build a dedicated integration for any single agent product yet, given that two of the four major launches in this space were already discontinued or reworked within roughly a year. The infrastructure decisions worth making now are the general ones (clean markup, accurate on-page data) rather than bets on a specific, still-unsettled product surviving.

Agentic browsers are a real, fast-moving category, not hype — real companies shipped real products with real users. But the discontinuation rate so far is a legitimate reason to treat this as an area to monitor closely rather than one to build a dedicated strategy around just yet.

A reasonable guess at why the shutdowns happened so fast

None of the four companies have fully explained their specific reasoning publicly, but the pattern across Operator’s folding into ChatGPT Agent and Mariner’s folding into Gemini Agent is telling on its own: both got absorbed into a broader, more general agent capability rather than being killed outright. That suggests these weren’t failures of the underlying idea so much as products that launched as separate, narrowly-scoped experiments before their parent companies had figured out where agentic browsing actually belonged in their overall product lineup. Consolidation into a single, broader “agent” feature — rather than a standalone browser or tool — looks like the direction both Google and OpenAI have converged on, even if Perplexity’s more standalone Comet browser suggests it isn’t the only viable structure.