# Retrieval Teaches Agent Sight

The first people to understand agents were the people who had already learned to search the internet as a medium.

Searching the internet well means learning what shape a fact would have if it existed outside your head.

A memory might be a post, a reply, a screenshot, a PDF, a filename, a quote inside someone else's summary, a cached page, a newsletter link, a handle that has since changed, an image with no text, or a phrase the source never used because the phrase was yours. Search teaches this through embarrassment. You ask the web for the concept and get advice. You ask for the exact phrase and get nothing. You ask for the document the situation would have forced someone to make, and the world begins to answer.

That skill looks like clutter from outside. Bookmarks everywhere. Files in strange folders. Half-remembered domains. Raw URLs. Screenshots with no OCR. The shape is messier than a bookshelf because the web's memory is arranged by everybody at once. A serious searcher learns to carry a model of that mess: which platform keeps which objects, which names change, which words belong to the source rather than to the searcher, which absence is evidence, and which absence is only a bad query.

Then agents started searching in public.

That changed the interface. A model's answer arrives after many hidden operations, but a tool call exposes one small, valuable thing: where the model thinks reality might be. A search query is the agent's current hypothesis about the outside world. It names the object class, the source prior, the vocabulary prior, the time horizon, and the failure it is trying to avoid. It is a confession made before the prose becomes confident.

This is why visible search feels like sight. Neural attention remains inside the box. Tool attention touches the world. The query is where internal compression spends itself against an external graph.

I watched a small version of this in an AI chat. Asked to add narrative around a local chess tournament, the agent searched by named account, adjacent accounts, tournament language, chess and puzzle motifs, and exact "mate in 1" phrases. The useful part was the triangulation. The agent was asking: is the missing object a post from this person, a post from someone near him, a tournament artifact, a chess-puzzle memory, or an exact phrase?

A retrieval-heavy human can read that immediately. The account restriction is a source bet. The tournament terms are a topic bet. The motif terms are a memory bet. The quote marks are a phrase-existence bet. The ORs are an aperture-control bet. If the trace misses, the correction is obvious: try the older handle, search "OTB" or "club" instead of tournament, leave the platform, search images, add date windows, chase replies, look in the local archive, or ask whether the remembered thing was never public.

That is the new operator skill. Read the path while it is still cheap to change.

This also explains why search-native companies see a different AI frontier.

Perplexity is the clean company-level instance. Aravind Srinivas came through AI research and built Perplexity around search: an answer engine where the model has to ground itself through sources. That origin matters. A model-first lab tends to ask how a model can use tools. A search-first company tends to ask how intelligence should make contact with the world.

Those questions produce different products. Answer boxes, citations, follow-up queries, browser sessions, workspace search, enterprise context, and personal data connectors are part of one expansion. The company keeps pushing on the same boundary: where does the agent look, what does it trust, and how does the human correct the looking?

The advantage lives deeper than citations or a copied interface. It lives in the error distribution. Search companies are forced to learn when sources are stale, when snippets lie, when popularity looks like authority, when the user's words are the wrong words, when a question should become a browser task, when a browser task should become an agent workflow, and when the model should admit that the world has not answered.

Other founders search different media. I read Musk's native medium as physics plus public coordination: factories, rockets, cars, satellites, a social graph, the state of human attention. That produces a different agent instinct: connect the model to the public conversation and then to machines. Perplexity's native medium is web contact. Google's was intent at planetary scale. Hari's is a private graph and an inbox-shaped boundary. The founder's search medium becomes the company's sense organs.

That makes agent trace design more important than it looks. Hiding every tool call can make the interface feel smooth while removing the human's best calibration channel. Showing every raw call can turn sight into exhaust. The right display gives the human the bet: what the agent believes it is looking for, where it looked, what it ruled out, and how the human can redirect the next reach.

The correction can be tiny. "Search this account too." "Use the older handle." "That phrase is my summary, not the source's wording." "This would be in an image." "Look for the document type, not the topic." These are edits to the agent's sight before they become edits to the answer.

Many agent errors begin before reasoning. The model reasons from the wrong retrieved world, or from the right world found through the wrong path, or from a source that ranked high because the query pointed at popularity instead of proof. By the time the final prose arrives, the search mistake has already become context. A human who can read the trace can stop the mistake upstream.

The black box stays black. When it reaches for the world, the reach has shape.

Searchers learned to see shape in disorder. Agents now externalize that shape at machine speed. The person who can read it can teach the creature where to look before teaching it what to say.