The browser wasn't the product either.

Ran three tasks through a browser agent this week: verifying my analytics pixel, checking a corporate site’s live deployment state, and walking a client’s checkout flow end to end. All three worked. None of them fired analytics. The agent behaves like a user, sees the page like a user, clicks like a user, but the tracking stack the industry has spent fifteen years building sees nothing.

Which is why the news that OpenAI is shutting down Atlas on August 9th matters less than it looks. Atlas was the standalone browser bet, launched last October, and it lasted nine months. What isn’t ending is the agent capability. That capability is moving directly into ChatGPT, which now drives web pages on the user’s behalf from inside the app most consumers already open.

Anthropic’s Chrome extension does the same thing from the other end. Every serious frontier lab now has a way for its model to see, click, and type on a live web page. It works. I’ve been living with it several days this month.

Perplexity’s Comet is the honest counter-example, near 18 million monthly active users after going free in March and running an agentic completion rate above 70%. It sharpens the point rather than undermining it, because Comet survives by explicitly not trying to replace Chrome, positioning itself instead as a research destination people visit for a specific job rather than a general-purpose browser they live inside all day. Atlas made the opposite bet and lost inside a year. The browsers that hold ground as standalone products carve off a job; the agent capability itself, which is the valuable part, ends up moving everywhere else.

Three implications the board should be discussing on Monday.

First, your funnel measurement is quietly losing coverage. When the “user” is an agent, the pixel doesn’t fire, the cookie doesn’t set, the consent banner is dismissed by script. The CRO team is not wrong. The number just no longer describes the audience.

Second, the “browser as product” bet failed publicly in nine months. The “agent as feature inside the surface people already use” bet is winning. Any product roadmap that assumes users switch destinations to access AI features is buying the same story Atlas just told.

Third, the pricing model for these agent runs is per-token consumed, not per-seat licensed. The governance gap I wrote about two weeks ago now applies to your own team’s agent usage too, not just vendor consumption. Every agent task run against your site is someone’s token bill, and every agent task run from your team is your token bill.

Boards spent the last decade optimising for click-through rate. In 2026, half the meaningful traffic doesn’t click. It executes.