I'm building a database where every AI agent writes to its own branch first. Looking for people to build it with.

An agent with write access to a production CRM deleted all our lead lists. It needed write access to do its job, so "make it read-only" wasn't an answer. So I'm building a database where every agent gets its own branch: a private copy of the whole database, ready in well under a millisecond. The agent reads and writes freely there. Then you see exactly what it changed, row by row with before and after values, and either merge it or throw it away. I asked people running agents what they'd need to trust this. Their answers shaped most of it: - Merges are three-way per row. If production changed a row the agent also changed, nothing merges. - The database records what the agent read. If that data changed while the agent worked, the merge is held, because the agent decided on facts that are gone. - You write SQL checks that run inside the merge, against the result. Catches two changes that are each fine alone but wrong together. - Rules decide what merges by itself: max rows, no deletes, tables or columns that always need a person. Limits count across the whole run, so splitting a delete into small pieces doesn't get around them. - Emails and webhooks queued in a branch only fire if it merges. - The agent's own claims are never trusted. The database records, a person sets the rules. It speaks the Postgres protocol, so psql and normal drivers work. Honest status: it works, it's public, and nobody outside me has run a real agent against it yet. I'm building it alone and I'm looking for: People whose agents write to real data (CRM, orders, tickets) who'd try it with me and tell me what breaks. Rust developers who want to work on a database engine: merge review, Postgres compatibility, tracking what reads a changed field. People who think the approach is wrong. I want that too. Comment or DM. Repo and a 2 minute demo in the comments, per the rules.