Agentic browsers in 2026: what actually shipped

The standalone bet is dying

On August 9, 2026, ChatGPT Atlas stops working. OpenAI confirmed the shutdown date less than a year after Atlas launched on macOS in October 2025 — a run so short that users are being told to manually export bookmarks before access ends, since nothing carries over automatically. The agentic capabilities Atlas tested don't disappear; TechCrunch reports OpenAI is redistributing them into the ChatGPT desktop app, a Chrome extension, ChatGPT Work, and Codex. The framing that stuck: the browser was a feature OpenAI shipped standalone to learn fast, not a product it intended to keep running forever.

Atlas isn't the only standalone bet that ended in an acquihire-shaped exit. The Browser Company — the startup behind Dia and, before it, Arc — agreed in September 2025 to sell itself to Atlassian for $610 million in cash, with the deal expected to close in Atlassian's second fiscal quarter of 2026. Atlassian says it'll run the team independently, but the outcome is the same shape as Atlas's: a browser built specifically to be the agentic front door got folded into a company that already has one (in this case, a work-software incumbent), rather than growing into an independent platform.

Two different endings, one pattern: building a whole new browser around an agent turns out to be a hard way to win distribution, even when the underlying agent capability is real.

What's actually shipping

Set the deaths aside and three products are shipping agentic browsing features right now, each betting on a different distribution wedge.

Perplexity Comet took the opposite path from Atlas — free and everywhere, instead of premium and standalone-only. It launched in July 2025 to Perplexity Max subscribers at $200 a month, then went free worldwide that October, with Perplexity CEO Aravind Srinivas confirming on CNBC that free users get the core assistant with rate limits. Comet reached all four major platforms — Mac, Windows, Android, and iOS — when the iOS build shipped in March 2026, the last piece of cross-platform coverage. It's Chromium-based, so most Chrome extensions still work; the agentic layer sits on top as a sidebar assistant that searches, organizes tabs, drafts email, and completes multi-step tasks like checkout flows.

Microsoft Edge's Copilot Mode takes the incumbent-browser path: don't build a new app, retrofit the one people already have installed. Microsoft introduced it as an opt-in, free-for-a-limited-time mode on July 28, 2025, on Edge for Windows and Mac, layering an assistant that reads open tabs and answers questions on top of ordinary browsing. The agentic piece — Copilot Actions — lets the assistant click, fill forms, and complete flows across tabs with explicit consent, aimed at repetitive tasks like price comparisons and re-filling the same form fields. Microsoft pushed the enterprise version further in late 2025, putting agentic browsing for Edge for Business into limited preview under IT policy control — approved-site allowlists, admin-set task boundaries — with general availability targeted for the second half of 2026.

Brave Leo's agentic browsing takes the incumbent-browser path too, but with a narrower, more inspectable design. It entered early access across all Brave release channels in May 2026, running multi-step tasks in an isolated browser profile rather than a hidden background process — the agent's session plays out in an open tab the user can watch. Brave's own description of the safety model is specific: users can inspect and pause a running session, and the agent can't delete its own session logs. That's a narrower agentic surface than Comet or Copilot Actions ship today, but the constraint is the point for a company whose whole brand is privacy-first defaults.

The consistent design problem

Whichever wedge wins distribution, the products converge on the same open question: how do you supervise an agent operating a browser on your behalf, on interfaces it doesn't control and wasn't designed to be operated by software? Brave's answer is visibility by default — an open tab, not a hidden sidebar — plus pause and inspect controls the user can reach mid-task. Microsoft's answer for the enterprise version is policy scoping before the session starts: IT sets which sites and actions are in bounds, so the supervision problem gets pushed upstream to configuration rather than left to per-action prompts. Neither approach is obviously right yet; both are early enough that the failure modes haven't fully surfaced.

That's the same territory covered in enterprise agent governance — who gets to approve what an agent touches — except here the "tool" is a full browser session with all the ambient risk of a live checkout page or a logged-in inbox. It's a harder supervision surface than a typed API call, because there's no schema to gate against, just pixels and clicks.

What to watch

Comet is the interesting holdout: it's still a standalone app, still growing, and still free, which cuts against the "browser is a feature, not a destination" reading that Atlas and Dia's endings support. Whether that holds depends on whether Perplexity can keep subsidizing free distribution — it raised fresh funding this year, but a standalone browser is an expensive thing to give away forever. On the incumbent side, the number worth tracking is Edge for Business's move from limited preview to general availability in the second half of 2026 — that's the point where agentic browsing under IT policy stops being an experiment and starts being something enterprises actually deploy at scale.

The tracker keeps a running read on where agentic browsing stands as it develops.