Anthropic’s Computer Use, OpenAI’s Operator, Perplexity’s Comet — three different bets on the same idea. The browser, as we know it, is on borrowed time. The agentic web isn’t a feature; it’s a paradigm shift the size of the desktop-to-mobile move. Here’s where it actually goes.

What “agentic web” means

Instead of a human clicking through pages, an AI agent navigates, reads, fills forms, makes decisions and reports back. The interface becomes a goal expressed in language; the browsing happens in the background. Conceptually obvious; technically only just shipping in 2026.

What’s already shipping

  • Anthropic Computer Use — Claude can drive a desktop, open apps, click, type, screenshot, recover from errors. Available to API users.
  • OpenAI Operator — GPT-driven web agent, runs in cloud sandbox, handles real tasks like booking flights or filling forms.
  • Perplexity Comet — agentic browser, search-first, integrates web actions into the search interaction.
  • Browser Use, browserbase — open-source frameworks letting any developer build agentic browsing into a product.

None are mainstream yet. All are good enough for narrow tasks, fragile for broad ones.

What breaks first

The places agentic browsing kills entire user flows in the next 24 months:

  1. Comparison shopping — “Find me the cheapest flight from RUH to LHR with one stop, by next Tuesday.” The agent does the 30 minutes of clicking; you accept or reject the result.
  2. Form filling — bureaucratic web forms (insurance quotes, government applications) all become single-prompt operations.
  3. Research synthesis — “Read the last 10 pieces of coverage on company X and summarise the financial signals.” Already deployed in production by hedge funds.
  4. Multi-tab orchestration — anything that requires juggling 20 tabs becomes a single agent task.

What this does to web business models

The implications are uncomfortable for current web businesses:

  • Display ads collapse — agents don’t see ads. Bots strip them out.
  • SEO traffic shrinks — agents pull answers, not page visits.
  • Paywalls get harder — agents either get blocked, or services have to negotiate machine-readable access tiers.
  • Brand equity matters more — when the user never sees your homepage, your brand has to live in places agents do see (schema, llms.txt, structured data).

What this builds

New surfaces emerge:

  • Agent-readable APIs — sites that publish clean MCP servers or structured data outpace those that don’t
  • Agent-first commerce — products described in language patterns agents understand and trust
  • Brand-as-citation — being the source AI agents quote becomes the new homepage rank

The 2027 forecast

By 2027, ~30% of consumer web tasks happen via agent. Power users hit 70%. The browser shrinks to a viewer for things humans still want to read directly — articles, video, social. Everything transactional moves to agent. MCP-style structured access becomes table stakes for any company that wants traffic.

What founders should do today

  1. Publish a clean llms.txt at your domain root
  2. Audit your schema markup — make sure entities, products, articles are machine-parseable
  3. If you have an API, ship an MCP server alongside it
  4. Stop optimising solely for human eyeballs; start optimising for agentic citation

The honest part

The browser isn’t dying tomorrow. Humans still want to read, watch, scroll, fall in love with brands. But the share of web tasks done agentically rises every quarter from here. The companies that prepare now win the next decade; the ones that don’t get scraped by agents and get nothing back.

Building for the agentic web? Trade notes.