Agents do the work. You just decide. Every step is on the record.

Traditional strata management hands one company your money, your maintenance, and your meetings, then asks you to trust that it is all being done well. GoodStrata splits that job in three. Agents (software, not staff) do the operational work the moment it needs doing: issuing levies, matching payments, following up arrears, triaging repairs and dispatching contractors, sending meeting notices and drafting minutes. You and your committee do the one part that was always meant to be yours: you decide. Approve the budget. Approve the quote. Vote at the AGM. And because every action an agent takes is written to a log that cannot be edited or deleted, you are never asked to simply trust. You can check.

An agent can prepare a decision. It can never make one. When something needs a person (adopting the annual budget, signing off a repair over the building's threshold, commencing debt recovery against an owner, or anything the Act reserves for the committee or a vote), the agent stops and opens an item in the decisions inbox. It arrives with the evidence already attached: the figures, the quote, the photo, the relevant clause, the reason it is being proposed. You open it on your phone, read what is proposed and why, and choose. When you approve, code carries out exactly what was proposed: not the agent's later interpretation of it, the precise action you saw and signed off. Until a person in the building acts, nothing about your money or your property changes.

The part of strata that has to be right (the money) is the part we never let the AI touch. Levy apportionment, penalty interest, payment matching, fund splits, and vote weighting are ordinary, deterministic code with exhaustive tests. The model writes the covering words; the engine computes the numbers and appends them.

In our launch run, an agent chased Lot 7's arrears with a figure correct to the cent ($1,153.85 overdue plus $14.86 penalty interest), and it would have stayed correct even if the model had tried to invent a number, because GoodStrata appends a code-generated statement block to every money email. Put plainly: if every AI on earth were swapped for a random number generator tomorrow, your levies would still sum to the budget exactly.

Everything an agent, a person, or the system does is written to an append-only event log: a running record that a database rule prevents anyone from editing or deleting after the fact. Each entry names who did it (a person, a named agent, or the system) and links back to what triggered it, so you can follow any outcome all the way to its cause. Human actions and agent actions are marked differently everywhere they appear, so the machine's hand is always distinguishable from a person's. Open the Agents tab and you can read every individual step a model took, down to each tool it called. That is what "auditable" means here: not a monthly statement you are asked to believe, but the whole manager's working, on the record, yours to read.

Here is roughly a month of management, the way it actually ran on our seeded twelve-lot walk-up, on one laptop, on a local model, with exactly one human tap. Every item below is a real event from that run, and every one of them is on the audit log.