Why not just point Claude Code or Codex at our company files?

You can. For one person on one project it works well. For a whole business it runs into five limits, and none of them is the model's intelligence.

Leo runs on the same class of model. It doesn't make the AI smarter. It gives it better material to answer from.

The short answer

	A coding agent on your files	Leo
Where an answer starts	Cold, every session	From a picture already joined up
What it reads	Raw documents, as many as fit	A few finished pages about the question
Who it learned from	Whoever wrote things down	Everyone who does the work
How sure it is	It doesn't say	Confirmed by a person, or only seen in evidence
When it doesn't know	It often fills the gap	It says nobody has taught it that yet
Your words for things	It needs the words the files use	It learns the words your people use

Every model works through a window

Claude Code, Codex and Leo all answer from what is loaded in front of the model at that moment. Nobody escapes that. The window is large, and it is still far smaller than a business.

So what matters is what goes into it.

A coding agent decides as it goes. It searches your folders, opens what looks relevant and builds a picture on the spot. Ask the same question tomorrow and it builds the picture again, possibly from different files. If the material is messy, the answer depends on what it happened to open.

Leo does that work before anybody asks. Each time somebody teaches it something, it rejoins what everybody has taught into a page for each piece of work, each system, each part of the business and each handoff between them. A question reads a few of those pages. It never has to dig through the transcripts, emails and documents they came from.

Documents say what was written. People know how the work is done

Point an agent at a shared drive and it learns the business as it was documented. That is the policy, the template and the process map from two years ago.

How the work really runs is in people's heads. The step everyone skips. The spreadsheet that is the real system. The person everybody asks.

Leo learns that by conversation, from the people who do the work, in their own words. Then it plays it back so they can correct it. That is where most of the picture comes from, and it is the part a folder of files can't give you.

You can check every answer

Every page in Leo carries two things: where it came from, and whether a person has confirmed it or it has only been seen in evidence.

When Leo answers, it names the page each part came from. When it has no page on something, it says nobody has taught it that yet. It is told never to fill the gap from general knowledge about businesses like yours, because you couldn't tell an invented answer from a real one.

A coding agent has no such standard unless you build one, and keep it up.

It learns the words your business uses

Search is where a company's knowledge quietly fails. One person calls it debtor follow-up. Another asks about chasing late invoices. An ordinary search treats those as unrelated, and reports nothing found about something the business knows perfectly well.

Leo is built so the page is found either way, and so it gets better at this the more your people ask. The words one team uses for a piece of work become a way for everybody to find it.

While you teach, it does the same in reverse. Use a new name for work Leo already has a page on and it asks whether they are the same thing, so one piece of work stays one page however many names it has.

One brain, not one per person

A coding agent's memory belongs to whoever is running it. Twenty people with twenty agents hold twenty partial pictures, and none of them knows what the others learned.

Leo is one picture that everyone teaches and everyone asks. What you say about yourself stays private to you. What you teach about the work is there for the whole business. Your material stays inside your own boundary: its own database, its own repository, and no outside search service.

You don't have to choose

This isn't Leo against the assistant you already use.

Leo connects to Claude and ChatGPT, so you can ask it from the chat you already have open. When you approve a build, Leo hands the work to a coding agent, with the context that agent would otherwise have had to rediscover.

The agent is the hands. Leo is what it knows about your business before it starts.

When a coding agent alone is enough

• You're one person, and the knowledge is in your own head.
• The work is one codebase or one project, not how a company runs.
• You're happy to check the answers yourself.

Past that, the question stops being which model is smartest. It becomes what the model has been taught, by whom, and whether you can check it.