The stale-fact tax

Two of us run about 40 projects - me, and a team of AI agents. I make the decisions (price, terms, who a product is for, what comes first) and the agents do most of the work. Building, writing, checking, deploying. I review key decisions, tests, architecture.

It works. Mostly. But there's a tax on it nobody warned me about, and for months I paid it without noticing.

The tax

Here's how it goes.

I ask an agent to write a post for one of our products. It goes and reads around first - a plan doc, a pricing note, its own memory from last week. It writes a lovely post. With last month's price in it.

I spot it. I correct it. The agent apologises, goes back, re-reads, rewrites. Then it updates one note - but not the other three that say the same wrong thing. Next week a different agent reads one of those three.

Multiply that by 40 projects.

This isn't hypothetical. In one session this week, while planning content, the agent working with me found:

- a launch plan written a week earlier that was built around a pricing model we'd since dropped

- a funding application that the plan board said was due in a fortnight, and my notes said we'd parked

- a programme described as six weeks in one doc and four meetings on the live page

- its own memory of one product's pricing, carrying three superseded versions on top of each other

None of these were mistakes when they were written. Every one was true on the day. That's the problem - a note is a snapshot, and snapshots don't know they've gone stale.

And an agent never hesitates. A person reading an old doc might think "hang on, didn't we change that?" An agent just repeats it, confidently, and builds on top of it.

So the real cost isn't the wrong price. It's the walk back:

- the tokens spent writing the thing on a bad foundation

- the tokens spent correcting it (and every correction re-sends the whole conversation)

- the tokens spent finding and fixing the other copies - if anyone bothers

- and the most expensive one, my attention. I'm the only human in the loop. Every time I have to catch a stale fact, that's me doing the agent's job.

It got bad enough that the very first line of my instructions to every agent is now, in capitals, "do not state unchecked claims as live fact". That rule exists because an agent once told me an incident was still active based on a log that was five days old.

The Pause

We're in the middle of a push to launch everything that has legs. The rule on the board is literally "minimum work to get things out - be suspicious of anything that improves something already shippable".

So stopping to rebuild internal plumbing looks bad. I know how it looks. I'm ADHD - I cycle between things, and from the outside this looks like one more shiny detour.

It isn't, and here's why.

In a normal company, a stale decision costs you one confused colleague. In a business where agents do most of the work, every single piece of output - every post, every sales page, every line of code, every email - starts with an agent reading what we've decided. If that read is wrong, everything downstream of it is wrong, at machine speed, across every venture at once.

It's not one of the things. It's underneath all of the things.

Fixing it once fixes it for every project. That's the same reason I built shared services instead of rebuilding sign-in and email for every venture - build once, use many. This is that, for decisions.

What is HQ?

HQ is the place our agents go to find out what's true across the portfolio. It plugs into every agent session (it's an MCP server - a standard way of giving AI agents tools).

It didn't start as anything grand.

June - it began as "the librarian". Agents working in one venture kept needing facts about another one, so I gave them something to ask.

Mid June - renamed HQ, because it was already doing more than answering questions. It picked up the roadmap and story maps, our automation workflows, a live view of the servers, CI status, creating repos.

August - an audit found that 16 of our ventures had been built without the agents ever seeing our standards (security, accessibility, honest UX). The standards only loaded if you happened to be working in the right folder. So HQ started pushing the mission, the standards and our working rules into every session, everywhere.

September - our legal details (company name, registered office, VAT) kept getting copied into footers and privacy pages, and drifting. So they moved into HQ once, and everything reads them from there.

You can see the pattern. Every piece of HQ exists because something went wrong without it.

But underneath, it was still a pile of files. A JSON list of projects. A folder of markdown for working rules. A config file for the company. Better than scattered across 40 repos - but still copies, still snapshots, still no way to say "this one is current, that one is history".

This is the bit I paused for.

I got pissed off - as I am wont to do - so I went to fix it.

Everything now lives in one small database. Every decision has exactly one home, with a name like "this product's price" or "the portfolio's priorities". Each one records what was decided, who decided it, when, and why - plus every previous version.

The nicest part day to day: when I decide something, it's recorded on the spot. In a recent product review I just talked - answered questions in plain sentences - and each answer became a recorded decision with a one-line receipt back to me. The sales page edits after that were checked against those, not against anyone's memory.

The whole thing is one file, under a megabyte. About 44 projects, a couple of hundred decisions, 30-odd working rules. Not big. It doesn't need to be.

How we improve it

We don't design this stuff up front. One of our working rules is literally "don't build rules or process for problems you haven't seen yet".

So HQ grows by pain. Something goes wrong, we ask "what would have stopped that?", and if the answer is general enough, it goes into HQ for every project at once. The librarian, the standards push, the legal details, the decisions database - every one of those started as a specific annoyance.

What's next is the same. There are still stale copies sitting in older docs around the ventures, and they need sweeping out. The old files need retiring. And I want a simple read-only view so I can browse the decisions myself, without asking an agent.

Will something else go wrong? Of course. Then we'll fix that too.

The solution

If you're running agents in your business - or about to - here's what actually fixed it for us. None of it needs our tooling. It's mostly a way of thinking.

One home per decision. Not "the latest doc". One place, and everything else points to it. A copy is just a decision waiting to go out of date, so docs and notes now say "see this product's price" rather than writing the price down.

Bring the truth to the agent. You can't rely on an agent choosing to look something up - it won't, reliably. So every session starts with the decisions for its project already loaded. And when I change a decision in one session, every other open session is told on its next message: "this replaces anything you said or remembered before".

Show age, not dates. "Decided 9 days ago" makes an agent (and a person) ask whether it still holds. "2026-10-05" doesn't.

Write down what you haven't decided. A gap is an invitation - an agent missing a price will invent a plausible one. So a decision can be an open question, and agents see it as exactly that: do not fill in. One of these stopped a sales page from hinting at a price rise nobody had agreed.

Fail loudly. If the truth can't be read, the agent isn't quietly handed an old copy. It's told: you don't know, don't quote anything from memory, ask.

Record it when it's decided. Not at the end of the week, not "I'll update the doc later". In the same breath, with a receipt.

The point isn't a perfect system. It's that when an agent tells me something now, I can trust it's what we actually decided - and I can spend my attention on the work, not on catching the past.

So - where do your team's decisions actually live?