A few editions back I told you I had to leave Notion. I told you the system had outgrown a notebook, and that moving off it was the difference between a clever prototype and an actual system.
What I never told you was where I went.
So here it is. And the honest headline is this: the move was not the interesting part. What I learned building the thing underneath was.
Why a notebook can't be a foundation
Notion was never designed to be the engine room of a system. It is a place for humans to read and write. My agents were hammering it like a database — thousands of reads, thousands of writes — and it pushed back the way any notebook would. Rate limits. Slowness. The constant low-grade fear that the thing holding all my data was owned by someone else and could change under me.
You do not notice the foundation until you put weight on it. I had put a lot of weight on it.
Why Google
The choice of where to build was not really a choice. We are already a Google company — the email, the calendar, the documents all live there. Keeping the engine in the same place meant the data was not scattered across someone else's systems, and it came with the kind of security a business actually needs rather than the kind a side project gets away with.
The deciding factor, though, was the boring one: room to grow. I am the only user today. But a foundation you might one day put other people on is a very different thing from one built for an audience of one — and I would rather choose that now than rebuild it later.
Not one replacement. Two.
This is the part I got wrong in my head before I started. I assumed I needed a database. Singular. Swap the notebook for a proper filing cabinet and you are done.
What I actually needed was two different things doing two different jobs.
The first is Postgres, running on Google Cloud. (Postgres is a fifty-year-old idea done properly. Tables with strict shapes, every record with one address, and a system that flatly refuses to pretend a record exists when it does not. It is the filing cabinet that argues back. After watching an agent confidently quote a record that was not there, a thing that argues back was the entire point.) Postgres is the system of record. The single source of truth. If you want to know what is real — which partner, which conversation, which decision — you ask Postgres, and Postgres answers or admits it does not know. The second is Firestore, also on Google Cloud. (Firestore is built for speed and for messages, not for truth. It is where my agents pass notes to each other while they work, and where the dashboard grabs a fast copy so it does not have to bother the filing cabinet for every little thing.) Firestore is the fast layer. The messengers and the scratchpad.
The one rule that holds it together
Two stores creates an obvious danger. What happens when they disagree? I wrote the answer into the system as a flat rule: if Postgres and Firestore disagree, Postgres wins. Every time. One of them is allowed to be the truth. The other is only ever allowed to be fast. The moment you let the fast one start being believed, you are back to an agent quoting something that was never real. That single rule — one source of truth, everything else is a convenience — is more important than any agent I have built.
The boring numbers
Over a weekend in early May, 2,026 records moved across into ten tables in Postgres. A database in London, a proper schema, a real backend. It is not a good demo. Nobody claps for a migration. But it is the line between a thing that looks impressive in a screenshot and a thing you can actually trust to run.
The turn
Here is what four editions of agent stories were quietly building toward. I spent months thinking the system was the agents. The clever autonomous workers doing my job while I slept. That was the part I was proud of, the part I wanted to show people. The agents were never the hard part. The hard part was the floor they stand on. Get the structure wrong — the wrong store, no single source of truth, a notebook pretending to be an engine — and it does not matter how clever the agent is. It will confidently build on sand. Get the structure right, and a fairly ordinary agent becomes reliable, because everything it reads is true. I had it backwards for months. I was decorating the upstairs of a house with no foundations. You cannot put autonomy on a notebook!