I handed this project an entire computer, but that does not mean every AI process should get the run of the place. A bot reading random websites will eventually read bad instructions. A bot writing code will eventually write bad code. Neither one belongs beside passwords, account records, or anything that can move money.
The answer was to split the computer into three virtual machines. Think of them as three locked rooms with different jobs.
The three rooms
- The lab. This is where the messy work happens: reading sources, trying ideas, and running generated code in throwaway containers. It gets no account or database password.
- The office. This assigns work, watches API spending, checks handoffs, and runs the dashboard I use to see what everybody is doing.
- The vault. This keeps market data, trade records, risk decisions, the paper account, and backups. It does not surf the web or chat with models.
The workers can pass a request through a small controlled door. They cannot wander into the database. Experiments get limited memory, processor time, disk space, and internet access. When the test is over, the container gets thrown away.
A name does not come with a master key
I want the Librarian to remember what it already researched. I want the Skeptic to remember which ideas blew up. That does not mean they need the same browser, the same password, or unlimited access to the computer.
Memory and permission are separate. An agent can remember its job without being allowed to rewrite history or give itself more power.
This is for the day something screws up
The walls matter when something goes wrong. Old prices, a dead AI provider, a restarted database, a duplicate job, or a broken strategy should stop that one job. It should not poison everything else.
Let the research side get weird. Keep the money side boring.
Virtual machines are not magic safety boxes. They just make the blast smaller when something stupid happens, and they give me walls I can actually test.