You ask an agent a question that needs sources. It searches, gets ten results, reads all ten, then searches again with a phrase almost identical to the first one and reads those too. Nine minutes later it hands you a confident paragraph and two links, one of which does not say what the paragraph claims. Nothing errored. The trouble is that nothing was ever decided: at every turn it invented the strategy again from nothing, with no record of what it had already ruled out. Searching was a sentence in its instructions. It was never something the agent could pick up and use.

You never start at the highest magnification

A microscope carries three or four objectives on a rotating nosepiece, and there is one order in which people use them. The specimen goes on the stage and you begin at the weakest, the widest and dimmest, because at that power the whole sample sits in the field and you can see where the interesting part actually is. Then you turn the nosepiece one click. Then another, onto the spot you have centered, and only there does a single cell resolve into something you could describe out loud. Begin at the strongest and you spend twenty minutes hunting across a field the size of a pinhead.

Retrieval works the same way once it stops being one opaque call. Keyword search is the weakest objective: cheap, blunt, good for confirming that a library by that name exists and roughly where it gets written about. Semantic search is the next click in, narrowing to the theme rather than to the string. Reading a chunk is the strongest lens, and you use it on the two passages of documentation that bear on the sentence you are about to write, and on nothing else. Every result shapes the query after it, which is the part a fixed pipeline cannot do at all.

How much you read follows the question

The older arrangement fetches documents up front and pastes them into the prompt before the model has looked at anything. That is every sample in the box lifted onto the stage at once. The window fills with material whose relevance nobody has judged yet, and the model spends its attention weighing what it was handed instead of asking for what it needs. The price is the same whether the question was easy or hard, because it was set before the question arrived.

Granular tools invert that. The volume pulled in adapts to the difficulty of the task, because how deep to go becomes a decision made inside the loop rather than a setting chosen at build time. A stronger model can then spend its reasoning at run time on the path through the material, which is exactly the sort of judgment you wanted from it and could not get from a pipeline.

There is a trap on the way. Exposing one vague search call collapses discovery and targeted reading into the same thing, and an agent with one objective on the nosepiece cannot choose a magnification. That is the ordinary discipline of making a tool legible, applied to the one capability people most often ship as a single button.

The record is part of the instrument

Turning the nosepiece is only half of laboratory work. The other half gets written down as you go: what you were looking for, what the field showed, what is still unsettled, which sample to prepare next. Nobody holds that in their head, and nobody accepts a finding with no trail behind it.

Research toolsets make the same move, by turning the intermediate decisions into calls with typed arguments. Plan the next searches. Choose which query to actually run. Pull out the details that matter. Analyze whether the search is getting anywhere. The outputs name the hypothesis, the gaps in evidence, and what would count as progress, so a redundant query is visible as redundant before it is spent. After two inconclusive results the agent calls the progress analysis, records that a primary source is still missing, plans a query aimed at exactly that, and uses the extraction call to isolate the passage its conclusion rests on.

The gain is not cleverness. It is that the trajectory can be read afterwards by something other than the model that produced it, so an orchestrator can tell a search that found nothing from an answer that is genuinely grounded. The failure mode is bureaucracy: reasoning calls made because the schema asks for them, whose progress analysis never changes the next query. Scaffolding that nobody reads back is ritual. The strategy stops being improvised the moment it stops being an instruction.