Insight · October 15, 2026
Agents are only as good as
what they can read.
Every company wants one place where agents can read what the company knows. Then they open the folders.
01 · The shift
The first job is the content, not the model.
A recent post by Greg Isenberg, on what he heard from AI founders in San Francisco, described a pattern we see in every conversation. Companies want to pull their internal knowledge into one place and expose it to agents. Often that means a connector built on MCP, the open standard that lets an agent ask a system for information and actions.
The connector is the easy part. It moves whatever is there. It has no opinion about whether what is there is correct, current or readable. So the project stalls on a different problem than the one the team planned for. The agent is ready. The library is not.
02 · What the agent finds
Five versions of the same document.
Open any shared drive and look for the most used procedure. There is a final version. There is a final version two. There is a copy with a person name in the filename and a copy somebody saved to the desktop and uploaded later. Five files, one subject, no marker for which one is true.
A person asks a colleague and gets the answer. An agent does not ask. It retrieves the closest match and quotes it with full confidence. If the closest match is the wrong version, the answer is wrong and reads exactly like a right one.
Then there are the abandoned pages. A wiki page from years ago describes a process that was replaced. Nobody deleted it. And there are the PDFs nobody can search: scans, exports, files where the text is a picture. To an agent, a document it cannot read is the same as a document that does not exist.
The one line to keep
“An agent can only be as right as the page it reads first.”
03 · What to do first
Five steps before you connect anything.
01
Inventory
List what exists and where it lives. Count the copies of the ten documents people use every week. The count is usually a surprise.
02
Pick one source of truth
For each kind of document, choose the single location and the single version that counts. Everything else points to it.
03
Retire the rest
Archive duplicates and stale pages. Do not leave them where an agent can reach them. Deleting is not required. Hiding them from retrieval is.
04
Structure it
Give every document of a kind the same shape, the same fields and the same metadata. Turn scans into text. Make the file searchable.
05
Name an owner
Every document has a person who answers for it and a date it was last confirmed. Without an owner, the library decays again.
04 · Structure
Structure is what a machine can actually read.
Clean is not enough. A tidy folder of free form documents still forces the agent to guess where the answer sits. Structure removes the guessing. A fixed set of fields. A status that says current or retired. An owner. A version. A date. A heading pattern that is the same in every file of its kind.
Templates do most of this work. When every document of one kind starts from the same template, an agent knows where to look, and so does a new hire. People also write faster, because the shape is already decided. The template is the cheapest piece of structure a company can add.
None of this is glamorous. It is editorial work: deciding what is true, saying it once, in a shape that holds. It is also the work that decides whether the agent is useful on day one or embarrassing.
05 · Proof point
What this looks like at GE Aerospace.
Bttr. built FLIGHT DECK for GE Aerospace, the operations hub, together with its 149 templates. One home for how the work is done, and one template for each kind of document, so a file of a given kind has the same shape wherever it was written.
That is the property an agent needs from a library. Not more content. The same content, in one place, in a shape a reader can predict. The full story is in the FLIGHT DECK case study.
record standard work: changeover owner named person, confirmed on a date status current version one, with earlier versions archived format one template, fixed fields source one location, one link
A content record an agent can trust · one owner, one status, one shape
06 · Then connect it
With clean content, the connector finally pays off.
Once the library is in order, MCP or any other connector does what the slide said it would. The agent finds the current procedure, cites the right page and stops quoting the copy from three years ago.
Decide what the agent may read, and what it may change, before it starts. Most of the library should be read only. Then test it the way you would test any software: ask twenty real questions, check each answer against the document it should have used, and keep the score. That is the same discipline we describe in Evals Are the New Unit Test.
07 · Why it lasts
The content is the part you keep.
Models change every few months. A company that spent its effort on a clever setup around one model starts over each time. A company that spent it on clean, structured, owned content keeps every hour of that work, whichever model reads it next.
It also helps people. The same library that makes an agent accurate makes a new employee faster, and makes an audit shorter. Few AI projects pay back to the people in the building. This one does.
Closing
Fix the library before you hire the reader.
Pick one folder your agents will read first. Count the versions of its most used document. Choose one, archive the rest, and name the person who owns it. Then ask the agent a question about it.
Written by · October 15, 2026
Donny Smith · ECD, Founder, Bttr.
Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.
LinkedInA recent post by Greg Isenberg on what he heard from AI founders in San Francisco, paraphrased · Model Context Protocol · Bttr. work for GE Aerospace
Share this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.
Read before you connect