We ran an AI agent per client, all at once, to draft the knowledge Vinny needs for every building it works in. The result was roughly 75,000 lines of drafted knowledge in about five minutes. Here's why running the work in parallel changed everything, and what it took to trust what came out the other side.
The short version
The slow part of knowledge management was never the writing. It was getting the first draft out of someone's head, one conversation at a time.
The work for each client has no dependency on the work for any other. That makes it a parallel problem, not a sequential one.
Run every client at once and total time stops being the sum of the pipelines and becomes the length of the slowest one. Fifty clients cost roughly what one costs.
Nothing went live on its own. Every record came back as a draft for a human to approve, and search was scoped to known sources only.
Vinny is the AI teammate for property teams. It answers resident questions at 2am, triages maintenance requests, chases arrears, and carries an enquiry all the way through to a signed lease. Every bit of that rests on one thing: knowing the specifics of the building, and the organisation, it works for.
That knowledge is the hard part.
The bottleneck was never the writing
For every client, Vinny needs detailed, accurate information on organisation policies and property-specific edge cases. Which contractor handles the third-floor lift. Whether the pet policy differs by building. What day the bins go out at one scheme and not the next. The kind of thing that never lives cleanly in a PMS, because no PMS has a field for it.
It lives in the heads of the people who run the buildings.
Traditionally, getting it out meant back and forth with property managers, turning scattered notes and inbox archaeology into clean FAQ answers, each split into a record for a single topic. That's around ten to fifteen minutes to get one topic right. Across a client's full portfolio, it rolls up to months of cumulative effort. Multiply that by every client and you have a permanent tax on how fast Vinny can start being useful.
So we asked a different question: what if agents did the first draft?
The interesting part isn't the agent. It's the shape of the work
The obvious version of this is one agent working through the backlog, record by record, client by client. That works. It's also still a queue. If a pipeline takes four minutes and you have dozens of clients, you've automated the effort but kept the calendar time. Sequential work means total time is the sum of every task.
The knowledge work for one client, though, has no dependency on the knowledge work for another. Different buildings, different policies, different sources. Nothing shared, nothing to coordinate, no ordering to respect.
That makes it easy to run in parallel, and it changes the arithmetic completely. Run every client at once and total time stops being the sum of the pipelines and becomes the duration of the slowest one. Fifty clients cost roughly what one costs.
That is the whole reason this took five minutes instead of a week of compute.
What we built
An agentic harness that spins up a cloud-hosted agent per client, each with a single goal: confidently find and add information that improves the resident experience.
Every agent runs the same pipeline, start to finish:
Read the existing knowledge records
Restructure and categorise them
Consolidate duplicates and flag the gaps
Search approved sources to fill those gaps
Hand off to a review agent that checks the output for issues
The harness holds no shared state between agents. Each one is isolated, gets its own client context, and writes back independently. Adding a client adds an agent, not a wait.
Two constraints did the work on trust
Speed is only interesting if you can rely on what comes out of it. Two decisions mattered more than anything else in the architecture.
Nothing went live on its own. Every record came back into Vinny as a draft, for a human to approve before publishing. The agents produced candidates, not truth. A property manager still decides what Vinny is allowed to say.
Search was scoped, not open-ended. Agents could only draw on known sources: client websites and the portals where clients advertise their properties. Public property information and already-curated marketing content. We did not scrape for organisation policy, because organisation policy is exactly the category where a plausible-sounding guess is worse than no answer at all.
That second constraint is also why the review agent exists. When the sources are narrow, the failure mode isn't inventing facts, it's misreading marketing copy as policy. A second pass catches that before a human ever sees it.
The outcome
An estimated six months of human effort compressed into about five minutes, producing roughly 75,000 lines of drafted knowledge.
Every line of it a draft. None of it published without a human reading it first.
Where this goes next
This run was a one-off, but a strong signal of what's possible. We're now building the same capability into Vinny's platform: tools and workflows that help property managers draft and curate the knowledge base Vinny relies on, with the manager staying the editor throughout.
The lesson we're taking from it isn't really about agents. It's that a lot of operational work looks slow because it's been modelled as a queue when it was never actually sequential. The property sector is full of that shape. One building's renewals have nothing to do with another's. One scheme's compliance checks don't block the next.
The bottleneck in knowledge management was never the writing. It was getting the first draft out of someone's head, one conversation at a time.
This is the first post in Behind the Agent, notes from the team building Vinny. More on the engineering behind the harness soon.