Today VN Media started running on a team of AI agents.
Not a chatbot bolted onto a contact form — an operating model. Research, writing, editing, site operations and marketing, each owned by a specialised agent, all routed through a single coordinator, with me approving anything that touches the public. This post is the system’s first output: one agent drafted it, another audited it, and nothing you’re reading went live until I had read every word.
Why bother
I advise Australian small businesses on AI adoption. I wrote recently that the industry has moved past chatbots towards AI employees: software workers with defined roles and human oversight. The most honest proof I can offer is my own business running that way. If the pattern works, my clients get a version that has already survived contact with reality. If parts of it fail, I’ll publish that too.
There is a second reason. Running the operation myself is the fastest way to learn what agent orchestration actually costs in hours, dollars and mistakes. Reading about it doesn’t teach you where it breaks.
The shape of the system
Seven roles, launched in waves. Three are live today: a Coordinator that plans and routes all work, a Writer that drafts content, and an Editor that audits every draft against a voice guide and a list of banned patterns. A researcher, a webmaster, a marketing agent and an analyst join later, once the basics have earned their keep.
Two design choices matter more than any model or prompt. First, agents never talk to each other directly; everything routes through the Coordinator, so there is always one place to look when something goes wrong. Second, memory is plain files in a git repository: every rule and every decision is written down, and every agent action is a commit that can be reviewed and reverted. An agent team is only as good as what it writes down.
The guardrails came first
Before any agent did its first job, I wrote the rules it can never break. Every action is classified by risk. Reading runs freely. Writing to private repositories runs with logging. Anything public requires my explicit approval, every single time: publishing, posting, sending, deploying, deleting. No lane gets automated until it has earned trust, and none has yet.
Duties are separated by design. The Editor that audits content never publishes it. The Writer never deploys. The future Webmaster will never write content. If you have worked in a regulated industry you will recognise the maker/checker pattern, and it turns out to apply to software colleagues just as well as to human ones.
What I don’t know yet
The honest numbers as of today: zero posts had been through this pipeline before the one you’re reading, zero newsletter subscribers, and a setup budget of six to eight hours with an ongoing cap of five hours a week. Whether that cap holds, and whether my voice survives being delegated to a Writer agent, I genuinely don’t know. That is the point of doing it in public.
Following along
I’ll document this build as it happens, with real numbers where I have them. The next entry in the series will likely cover the first thing that breaks. If you’re weighing up what AI agents could do inside your own business, or you’re building something similar, I’m happy to compare notes — no pitch, just a conversation.
Ready to put this into practice?
VN Media Solutions works with Australian businesses to implement AI agents, automate manual processes, and build the systems that make it all work.
Start a Conversation