Browsing Is Becoming Manual Labor

There is a particular moment when a technology becomes native. Not when everyone adopts it, but when the people who use it heavily stop understanding why they ever did things the old way. For a growing group of people, browsing the web is approaching that point. Opening Google, clicking through results, navigating a site, setting filters, scanning pages and manually comparing options increasingly feels like doing work the computer should be doing for you.

This is not yet a universal experience. AI adoption is uneven, and plenty of people still use LLMs as slightly better search engines, if they use them at all. But once you become accustomed to agents, public datasets and tools exposed through protocols like MCP, your tolerance for traditional browsing changes remarkably quickly. The shift is subtle and then obvious. You open a website, see 400 results and a row of filters, and think: I have to do this myself?

The reason is not simply that LLMs know more. It is that they can create filters where filters could never practically exist. Traditional software can filter a housing search by bedrooms, price and square footage. An LLM can search for “apartments with lots of natural light, a real kitchen and a layout that does not make working from home depressing.” A product database can filter televisions by size and resolution. An LLM can look for “the best television for a bright room where I mostly watch sports and do not care about smart features.” These are useful constraints encoded in descriptions, reviews and other unstructured data that traditional interfaces were never designed to query.

Reviews make this capability even more valuable. A specifications page tells you what a product is supposed to do. Reviews tell you how it fails. An AI can read across hundreds of them and notice that the laptop with excellent benchmarks has distracting fan noise, that the apartment building has recurring elevator problems, or that the software product is easy to purchase but unusually difficult to cancel. These downsides were always technically discoverable, but only by opening pages, reading comments and identifying patterns manually. An agent can surface the warning before you make the wrong choice. It does not merely find the options that match your criteria. It can explain why the most obvious option may not be the best one.

That is the deeper transition. The internet already contains an extraordinary amount of the information we need. The bottleneck has been translating human intent into whatever taxonomy a website happened to build. Dropdown menus were a compromise between what humans wanted and what databases could understand. LLMs increasingly remove that compromise. The filter can now be whatever you can describe, including the qualities and hidden drawbacks that no product manager thought to turn into a checkbox.

Agents take this another step. Instead of asking an LLM to summarize what it already knows, we can give it access to current datasets and tools. MCP makes more of those systems legible to models. The result is something closer to delegating research than performing search. The agent retrieves, reads, compares, calculates, filters and warns. You specify the outcome.

Once you work this way regularly, pagination starts to look almost absurd. If you are scanning 80 results looking for the six that satisfy a combination of constraints, you are effectively executing a database query with your eyes. The machine has the data. It increasingly understands the criteria. Why are you still doing the filtering?

This is an IYKYK transition because technological habits do not change uniformly. There will be a period when two people can sit next to each other trying to answer the same question and experience completely different amounts of friction. One spends twenty minutes searching, opening tabs, reading reviews and comparing. The other describes the problem, lets an agent work through the sources, and spends two minutes inspecting the result. Both are “using the internet,” but increasingly they are using different interfaces to it.

There is one important exception: taste. We still browse when we do not yet know how to describe what we want. Sometimes the point of looking through clothes, furniture, architecture, photography or places is precisely to encounter something that produces a reaction. Taste is difficult to outsource because browsing itself generates information about our preferences. You see something and think, that. No prompt could have specified it beforehand because you did not know you wanted it.

This may become the dividing line. We browse to discover what we like. We use LLMs when we can describe what we want.

For everything in between, browsing increasingly feels like manual labor. Search engines made finding the right page dramatically faster. Agents are beginning to question why we need to inspect every page ourselves. Once that behavior becomes habitual, the inefficiency is difficult to unsee. Clicking through ten pages is no slower than it was ten years ago. It just feels slower because some people have already discovered the elevator.