Answer
How do you give an agent the right material without giving it everything?
Give it the means to find things and the specific items the first step needs. Supply on demand, not in advance.
Give it a small starting set for the first step plus a way to fetch the rest when needed. Loading everything in advance degrades the run and leaves you unable to tell which material was doing the work.
The instinct to front-load comes from a reasonable fear: if the material is not there, the system will make something up. That fear is well founded and it is addressed by the ability to retrieve rather than by pre-loading, and the two have very different costs. Pre-loading pays the dilution cost on every run for material most runs do not use. Retrieval pays a round trip only when something is actually needed.
The pattern that works has three parts. A map: what exists and where, at a level of detail sufficient to choose. The specific items the first step requires, supplied directly. And a tool that fetches by name or by query. The map is what makes the fetch tool usable, and it is the part most often missing — an agent that can retrieve but does not know what is retrievable will either not fetch or fetch by guessing.
The second requirement is that the agent must state what it read. A run that says which documents it opened, and which it looked for and did not find, converts an invisible risk into a visible one. Without it, an answer built on two of the five relevant documents is indistinguishable from one built on all five, and the missing three leave no trace.
There is a class of material that should be supplied directly rather than fetched: the constraints and decisions that govern how the work is done. These are not references to be looked up when relevant, because the agent will not know they are relevant until it has already gone the other way. Small, durable, governing material goes in the front; large, situational, referenceable material goes behind a fetch.
The consequence for how a business keeps its files is worth stating. Retrieval works well against material that is organised, named meaningfully and current, and badly against a shared drive with three versions of everything and no convention. The work of making material fetchable is mostly ordinary information housekeeping, and businesses that have done it get noticeably better results from the same tools than businesses that have not.
An index and a fetch tool beat a folder, because the folder makes every document equally present and none of them prominent.
Siddharth Sharma, Context Theory
Related questions
What if the agent does not fetch something it needed?
That is the real failure of this pattern and it is why the completion condition should require a statement of what was consulted. The recovery is usually to make the map more specific rather than to supply more material, because the agent skipped the item through not recognising its relevance rather than through being unable to reach it.
Does this apply to a small project with only a few documents?
Less so. Where the whole relevant set is small enough to be present without dilution, supplying it directly is simpler and avoids a class of retrieval failure. The pattern earns its complexity when the material exceeds what should reasonably be held at once, which arrives sooner than most people expect but not immediately.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Visibility lift in AI-generated answers from GEO methods | up to 40% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Aggarwal et al., "GEO: Generative Engine Optimization", Princeton / Georgia Tech / IIT Delhi / Allen Institute for AI — KDD 2024 · GEO-bench · 10,000 queries across 8 domains · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Pre-loading pays a dilution cost on every run for material most runs do not use, while retrieval pays a round trip only when an item is required, so the two approaches to the same fear have opposite cost profiles. | Comparing per-run context volume and outcome quality under pre-loading and on-demand retrieval for the same task set. |
| Software | A retrieval tool is unusable without a map of what exists, because an agent that can fetch but does not know what is fetchable will either not retrieve or retrieve by guessing at names. | Providing a fetch tool without an index and observing the retrieval attempts in the run log. |
| Response | An answer built on part of the relevant material is indistinguishable from one built on all of it unless the run states what was read and what was sought and not found, because omitted sources leave no trace in the output. | Comparing the documents named in a run's consulted list against the set relevant to the question. |
| Procurement | Small durable governing material — constraints and closed decisions — must be supplied directly rather than left to retrieval, because the agent cannot recognise its relevance until it has already proceeded contrary to it. | Placing a governing constraint behind a fetch tool and observing whether it is retrieved before the decision it governs. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one