Sokage AI, AI agents for property managers.

What this actually is

Sokage AI in one paragraph, for anyone who needs the short version

Sokage AI is a paid community and agent library for residential property managers who want to put AI to work on the parts of the job that are repetitive, deadline-driven and easy to get wrong. We build one AI agent for each of those jobs, publish it with its test record, and run the room where members configure it against their own portfolio and tell us where it breaks. A new agent arrives every month and members vote on which one gets built.

An agent here is not software and not a subscription to a tool. It is a set of instructions plus a tested procedure that runs inside your own Claude or ChatGPT account, reading a file your property management software already produced. Nothing is hosted by us. Nothing connects to your system. If we stopped existing tomorrow, every agent a member has built would keep working, because they are files rather than services.

Who this is built for

Residential property management firms roughly in the 50 to 500 door range. Big enough that owner statements, after-hours calls and late rent are a real recurring workload, small enough that there is no internal developer and no appetite for a six month software rollout.

It works above and below that range, with a caveat each way. Smaller firms get the same agents and a proportionally smaller saving. Larger firms get more from it but need to spend longer on configuration, because the thresholds, fee structures and escalation paths that the agents ask for get more complicated as the portfolio does.

The agents read exports from AppFolio, Buildium, DoorLoop, Rent Manager and Propertyware, and from plain spreadsheets. That list is not an integration list. It is simply every system we have seen produce a month-end report, and the point of working on the file rather than the system is that the platform stops mattering.

How an agent gets built here

Members vote on which job gets automated next, so the queue is decided by people doing the work rather than by us guessing. Once a job is chosen, the build runs the same way every time: write the procedure, decide what it must refuse to do, write the tests for the ways it could cause damage, run those tests across several models, fix what fails, and publish the result including the failures.

That last step is the one most tools skip. The owner statement agent shipped with all 159 of its test runs attached, across three different models, including 59 runs that were deliberate attempts to hijack it through its own data. None of those succeeded, and the count is published so the claim can be checked rather than believed.

Every result we publish is synthetic. No real portfolio, no real operator, no pilot. That is a limit worth stating plainly rather than dressing up: a guard holding against test data is evidence, not proof, and we would rather say so than let a number imply something it does not support.

How we work

The rules we hold ourselves to

• Every number has a source and a date, or it is marked unverified. We would rather publish a gap than a guess
• We do not invent client results or testimonials. What we publish instead is the artifact, the citation, and the failure log
• No agent gets a send button. Not as a setting either. A person releases every word that reaches an owner or a tenant
• Jurisdictional questions stay yours. Habitability thresholds, fee bases and legal standards are operator blanks, because we are not your lawyer or your assessor
• We publish what breaks, and credit whoever found it