What is a "company brain", and how do you build one agents can use?

An owner asked me to connect an agent to the company Google Drive and let it figure things out. The Drive went back to the company’s first year. It held three versions of the price list, two of them wrong, and a proposal template nobody had used in years. The agent found all of it and quoted the old prices with total confidence.

What is a company brain?

It is the written record of how your business works, organised so a person or an agent can find the current answer to a question without asking you. It holds facts, procedures, decisions, voice and client history. It is deliberately smaller than your file storage, because everything in it has been checked, dated and given an owner.

The name sounds grander than the thing. In most businesses it starts as one shared folder of plain documents. What makes it a brain is discipline: one page per topic, the current version only, and a rule that when reality changes, the page changes the same day.

What goes into a company brain?

Six kinds of page cover most businesses: the facts about what you sell and how, the procedures your team follows, the decisions you’ve made and why, how you write and speak, what you know about each client, and the limits on what an agent may do. Each kind needs its own owner and its own review rhythm.

Why can’t the agent just search all my files?

It can, and it will find three price lists and use the wrong one. Anthropic’s engineering team describes an agent’s context as a finite resource with diminishing returns, and advises finding the smallest possible set of high-signal information for the job. A pile of old drafts is low signal. A current page with a date and an owner is high signal.

The same guidance describes agents keeping structured notes outside the conversation and fetching information when they need it, rather than loading everything upfront. A company brain is that idea applied to a whole business: a short index the agent always sees, and the detail pulled only when a task calls for it.

How do you build one without stopping the business?

Build it from the questions people already ask you. For two weeks, capture every question that reaches you, group them by topic, and write one page per topic with the answer, the exceptions, the date and the owner. Then test an agent against those same questions and fix the pages wherever it gets an answer wrong.

1. For two weeks, write down every question someone brings you. A voice note on your phone is enough.
2. Group the questions by topic. The same topics will come up again and again.
3. Write one page per topic: the answer, the exceptions, the date it was last true, and who owns it.
4. Put the pages in one folder with a plain naming rule, and move old versions somewhere the agent cannot reach.
5. Turn any task you explain step by step into a written procedure, with one example of good output and one of bad.
6. Connect an agent read-only, ask it the questions from step one, and fix the page behind every wrong answer.

How do agents actually read it?

Through connections you approve. The Model Context Protocol is an open standard that lets an AI application connect to data sources such as files and databases. Anthropic’s Agent Skills package instructions and reference files that load only when a task needs them, so an agent can hold many procedures without reading all of them every time.

The Skills documentation describes the pattern well. Only each skill’s name and a short description sit in front of the agent all the time. The full instructions load when a request matches, and reference files load only when the instructions point to them. Organise your pages the same way, and a growing brain doesn’t slow the agent down.

What keeps a company brain from rotting?

• One named owner per page, a person rather than a department.
• A date on every page, so a reader can tell current from stale at a glance.
• Changes tied to events: a price change updates the page the same day it takes effect.
• When an agent gets something wrong, fix the page that misled it, then the answer.
• Archive old versions out of the agent’s reach instead of leaving them beside the current ones.

How do you know the company brain is working?

Keep the list of questions you captured in the first two weeks and ask them again every month, to the agent and to the newest person on the team. When both get the current answer without asking you, that topic is done. When either gets it wrong, the page behind it is the thing to fix. The count of questions that still reach you is the simplest measure you have.

What should never go in it?

Passwords, API keys and anything else that grants access. Regulated client data, which belongs in the system built to hold it. Personal details an agent’s job doesn’t need. The Skills documentation itself warns that anything an agent can read and act on can be misused, so the brain should hold what the work needs and nothing more.