You keep your agents current by treating business context like code: one master file, a dated changelog, and a rule that nothing changes in the business until the file changes first. Most founders update the offer and forget the file, which is why an agent quotes last quarter's price with total confidence. The whole job takes about twenty minutes a month.
Why does my AI agent keep quoting the wrong price?
I raised my membership price in March. For six weeks my own agent kept telling people the old number, politely and with confidence.
The price lived in five places: a prompt I wrote in January, a Notion page, a Google doc, a spreadsheet, and my head. I updated two of them. Every place I missed was a place an agent read from. That is drift, and it is not a model problem. It is a plumbing problem.
This is what context engineering for founders actually looks like in practice. Not prompt tricks. Plumbing. The model was fine the entire time. I was the broken part.
What is the single source of truth for my agents?
One file. Mine is called business-context.md. It lives in a GitHub repo I can edit from my phone, but Notion or a Google Doc works fine if that is where you actually live.
Two steps to set it up.
First, pick one home and move every fact about your business into it. Who you serve, what you sell, what it costs, what you say no to, how onboarding works, what happens when someone asks for a refund. Mine is about 1,800 words. Yours will be shorter than you think.
Second, change every agent to load from that file instead of carrying its own copy of the facts. This is the step people skip, and it is the only one that matters. A prompt that says our membership is $1,400 a month is a copy. A prompt that says see business-context.md for pricing is a pointer. A pointer stays true. A copy rots the day pricing changes.
If it is not in the file, the agent does not know it. That sentence is the whole system. I walk through the full file structure in the AI OS playbook if you want to see how mine is laid out.
How do I version control AI prompts and context without a big process?
A changelog at the top of the same file. Five lines per change. That is it.
Here is the exact format I use:
`
v14 - 2026-02-03 - Membership price changed from $1,400 to $1,500/month.
Affects: sales agent, onboarding agent, FAQ bot.
`
Version control AI prompts matters for one reason you will appreciate the first time something goes wrong: you can see what changed, when, and which agents cared. When an agent gives a weird answer, I open the changelog and find the edit that caused it. Last fall I shipped a tone rule that made every outbound email sound like a bank. Two lines in the changelog and I rolled it back in four minutes.
Tools: GitHub if you like diffs and history. A Notion database with date-stamped rows if you want something your whole team can read without a tutorial. Either one beats a folder of files called final-final-v3.
Bump the version every time. Never edit in place without a new row. If an agent read v12 yesterday and v13 today, you want to be able to prove that.
What is the checklist on the day something changes?
Six steps. Takes twenty minutes, and I run it before I announce anything.
- Open business-context.md before you touch the website.
- Edit the exact line and delete the old one. Two prices in a file means an agent picks the wrong one eventually, guaranteed.
- Add a changelog row: date, what changed, which agents care.
- Bump the version number.
- Push to every surface. Here is where it breaks: hardcoded prompts. I search every prompt for the old price and the old offer name. Last time I found the old price in nine places. Each one becomes a pointer to the file.
- Test with three questions that only have a right answer if the update landed.
That is AI agent knowledge base maintenance in a nutshell. It is a habit, not a project. First Monday of the month I do a fifteen-minute audit even when nothing changed, mostly to catch copies that crept back in while I was moving fast.
How do I know the update actually landed?
Ask the agent. Out loud, in a normal chat.
What does the membership cost? What is included? When does it start? If it hedges or gives you two answers, it is reading two versions, which means a copy survived somewhere.
Then ask the question that saves you: where did you get that answer? Good agents will name the file. When mine says business-context.md, v14, I am done.
One more thing I learned the annoying way. AI agent operating system updates and business context maintenance are two different jobs. I can upgrade the whole stack on a Sunday and still have every agent quoting March pricing on Monday. The model changes. The facts do not. Keep them separate in your head and you will stop blaming the tools.
FAQ
How often should I update AI agent context?
Whenever a fact changes, plus a fifteen-minute audit on the first Monday of the month. If nothing changed, the audit just confirms the file is still the one source everything reads from. That is what catches a copy someone pasted into a new automation three weeks ago.
Do I need a developer to do AI agent knowledge base maintenance?
No. If you can edit a Google Doc, you can run this. The only technical piece is pointing your agents at the file, and if you have already built agents you have done harder work than that. I am not a developer. I ran this from a GitHub repo on my phone in an airport.
What happens if I forget to update the file for a month?
Your agents get confidently wrong, and you find out from a customer. That is the expensive version. The cheap version is one changelog row and twenty minutes. I have paid both prices and I know which one I prefer.
If your pricing, offers, or processes changed this month, the people I love and admire do not need a better model. They need the file updated. That is the whole point, and it sets you up for success long after the next shiny tool shows up. I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step. You can get it at https://a.mastermindshq.business/ai-os-book.