Skip to content

Future Tech/interface

Agentic Browsing Replaces Most Manual Web Use

Claude for Chrome ships. OpenAI Operator ships. Browser-as-runtime for AI agents is here. The web stops being browsed by humans for most multi-step tasks.

// By 2028 · high confidence · disruption 9/10

Prediction

// 2028

By 2028, more than 40% of complex multi-step web tasks (booking, comparing, applying, researching) will be handled by AI agents, not humans manually browsing.

Confidencehigh
Disruption9/10

What dies

  • internet explorer

Who wins

  • Anthropic
  • OpenAI
  • Google

filed: 2026-05-24 · guptadeepak.com

The hook

Claude for Chrome shipped in 2025. Within weeks, users were delegating multi-step tasks (book a restaurant, compare three insurance quotes, apply to five jobs) to the agent. The user gets the result, not the journey.

Thesis. The browser stops being the UI and becomes the runtime. Humans express intent, agents execute. The pointing-and-clicking model that defined the web for 30 years becomes one of several patterns, not the dominant one for complex tasks.

The story

The current state

Browser-use, Multi-on, OpenAI Operator, Claude for Chrome, Perplexity Comet, and Browserbase-backed agents are all shipping. Anthropic's Computer Use API (October 2024) made browser automation a first-class capability for foundation models.

The inflection point

Two things converged in 2024 to 2025: vision-capable foundation models and browser-automation infrastructure. Playwright and Puppeteer existed for years, but the model that could see a page and decide what to click did not. Now it does.

The prediction

By 2028, more than 40% of complex multi-step web tasks are agent-mediated. Booking, comparing, applying, researching. Humans still browse for exploration and entertainment. They mostly stop browsing for routine task completion.

Who wins, who loses

Winners: foundation model labs, agent-runtime providers (Browserbase, AgentBay), and websites that publish agent-friendly affordances. Losers: any business model that depends on a human seeing the ads on the page, and SaaS products designed around human-only UI flows.

Timeline and risks

Adoption is faster than analysts projected. The risk is the cat-and-mouse with bot detection. Cloudflare and others are shipping agent-aware bot management. The equilibrium is likely an llms.txt-style permission file that sites publish to allow specific agents under specific terms.

First signals (verify today)

Claude for Chrome shipped 2025. OpenAI Operator in production. Browserbase at $40M+ ARR. Perplexity Comet browser shipping.

Key data points

  • Claude for Chrome launch: 2025
  • OpenAI Operator launch: 2025
  • Browserbase ARR: $40M+ in 2025
  • Anthropic Computer Use API: October 2024
  • Perplexity Comet browser launch: 2025

Contrarian angle

Every B2B SaaS company is about to discover that 40 to 60% of its traffic comes from agents, not humans. The product analytics, conversion funnel, and even the UI design choices made for human eyes will need rethinking. The CTR and SEO conversations will be replaced by agent-readability and agent-trust conversations.

The flip side

What this kills

The paired obituary in Tech Graveyard.

Read the obituary

FAQ

How does an agent fill out a form on my behalf?

Vision-capable LLMs read the rendered page, identify form fields, and use a controlled browser session to type, click, and submit. Anthropic's Computer Use API and OpenAI Operator are the canonical implementations.

Can websites detect agentic browsing?

Yes, with varying reliability. Behavioral fingerprints (mouse movement, typing cadence), TLS fingerprints, and request patterns all leak signal. The equilibrium is moving from detection-and-block to disclosure-and-policy.

What is Browserbase and why does it matter?

Browserbase runs headless browsers as a service, specifically tuned for AI agents. The category exists because every agent product needs a clean, stateful, secure browser session at scale. Browserbase is the infrastructure layer underneath.

Will SaaS pricing change when agents are the main user?

Per-seat pricing breaks when one human runs ten agents. Per-task or per-outcome pricing is the likely successor for many categories.

More from guptadeepak.com

Want the technical deep-dive behind this prediction?

Read the companion article

More from the interface desk.