The Browser is still a library
Why AI browsers keep reinventing tabs—and what must change next
The web is now a workspace, but your browser still thinks it's a library.
OpenAI launched ChatGPT Atlas for macOS worldwide on October 21, 2025, with a built-in ChatGPT sidebar and Agent Mode for Plus, Pro, and Business tiers. It's a real browser, with real tabs — and very real AI.
The capability is new. The metaphor isn't.
The Paradigm Problem
Tabs were invented for a web that behaved like a library. Today, the internet is a workspace — where we research, collaborate, and transact. Yet the headline moves of 2025 kept the same scaffolding.
- Perplexity launched Comet on July 9 — an agentic browser experience still organized around pages.
- Google added Gemini to Chrome in September. The AI assistant can summarize across tabs and is moving toward agentic actions—still within the browser frame.
- Microsoft's Mustafa Suleyman told The Verge that Edge will "evolve to become a true agentic browser you watch as it works." An evolution of today's UI, not a replacement.
- Atlassian acquired The Browser Company on September 4 for $610 million in cash. Dia and Arc join a team building "the browser for knowledge work." Same paradigm.
State of Play
OpenAI launched Operator on January 23, 2025, as a research preview for Pro users in the United States. A browser-using agent that handles tasks autonomously — clicking through websites in a browser window you can watch. Same tabs. Same assumption that you need to monitor its work.
Chrome holds 71.77 % global share as of September 2025. Page Views and URL destinations align with existing ad and analytics models. An agent that never shows a page breaks those incentives.
Alphabet's 10-K shows that more than 75 % of 2024 revenue came from online advertising. The interface serves the business model.
Why We’re Still Here
Economics explains part of it. Chrome's dominance drives Google Search usage, which generates three-quarters of Alphabet's revenue through advertising. Agent-based interaction — where you never visit a URL or open a tab — is harder to track, insert ads into, and measure engagement.
Compatibility explains more. Tabs are a universal contract among users, developers, and sites. Ripping them out risks breaking authentication flows, payment systems, and content gating.
Trust matters too. If the agent works off-screen, how do you verify what it did, what it touched, and what it cost?
What Agent-Native Interaction Should Mean
You state the goal. The system compiles a task graph — API calls when available, DOM automation as a fallback — runs it in the background, and returns a receipt you can verify.
Example:
Compare pricing and integrations for three project-management tools for a remote team, prioritizing SOC 2 and SSO.
The agent checks 12 sources, uses 2 vendor APIs, gathers 4 citations, and requests 1 approval for login.
Estimated cost: $0.12
Time to result: ~19 minutes
Output: a structured brief with links, caveats, and a one-click reproduction run.
Tabs become optional. The work is the product.
Where Tabs Still Win
Tabs work for parallel exploration, fine-grained control, accessibility workflows, and legal clarity.
You saw the page. You accepted the terms.
The claim isn’t “tabs are bad.”
It’s that they’re insufficient as the primary unit of work for complex tasks.
What’s Actually Missing
- Task graphs and orchestration — first-class objects for multi-step missions spanning APIs and pages, including retries, fallbacks, and rate-limit policy.
- Durable memory and context — cross-task recall without leaking private data; scoping and TTL rules. Atlas hints at this with “browser memories” but stops short of task-level memory.
- Permissioning and identity — scoped, expiring, user-approved capabilities (login, checkout, calendar write) with least privilege and safe revocation.
- Provenance and audit — signed step logs for every action; reproducible runs; source citations; scalable trust when supervision isn’t feasible.
- Metrics and economics — judge success by task completion, time-to-result, interrupt count, revert rate, and cost per task. Align publisher compensation with contributed value (API access, summary rights, link-backs).
Bottom Line
Leading teams proved they can embed formidable AI inside browsers in 2025.
But until the unit of interaction shifts from tab to task, we’ll keep shipping faster horses.
The agent-native primitives above make tabs optional — and often unnecessary.
But declaring independence from tabs is easier than designing what comes next.
The real work isn’t critique — it’s architecture.