I Run a Fleet of AI Bots and Became a Human Message Bus Between Them
By Jon Ng, Founder, Novatise · 4 min read · 30 October 2026
I run a fleet of AI bots across Novatise and our own sites in Singapore and Japan. Each one has a job: research, writing, checking prices, watching the sites, and so on.
For a while, the most annoying part of all this was not the AI. It was me.
The problem: I was the glue
Here's what a normal morning looked like:
- Bot A finds the data.
- I copy its output and paste it into bot B.
- Bot B writes the page.
- I copy that and paste it into bot C to check it.
- Bot C finds a problem, so I copy the note back to bot B.
Every bot was fast. The whole system was slow, because every hand-off went through one person copying and pasting between chat windows. Engineers have a name for the part of a system that passes messages between other parts: a message bus. I had become a human one.
It's a common trap. Adding a new agent feels like adding a worker. But if every agent only talks to you, each new one adds more copying, not less.
Why it happens
Most AI tools are built as one-to-one chats: you ask, it answers. Nothing tells bot B that bot A has finished, where the result is, or what to do with it. So the human fills the gap.
Three things were missing in my setup:
- A shared place for work. Each bot kept its output inside its own chat, where no other bot could see it.
- Clear roles. Some bots did a bit of everything, so it was never obvious who should pick up the next step.
- Hand-off instructions. Even when the output was available, the next bot didn't know what "done" looked like or what came next.
The fix: let the bots hand work to each other
This is how I'm restructuring the fleet. None of it needs fancy software, and you can apply the same ideas whether you use ChatGPT, Claude, n8n or your own scripts.
1. Use shared files as the meeting point
Instead of passing text through me, each bot writes its result to a shared file or folder that the next bot can read: a spreadsheet, a database table or a plain text file. The rule is simple: if it isn't in the shared place, it didn't happen.
This also gives you a record. When something goes wrong, you can open the file and see what each bot produced, instead of scrolling through five chat histories.
2. Give every bot one clear job title
I give each bot a short job title and a one-line description, like a team member:
- "Researcher: finds sources and saves them to the research file."
- "Writer: turns the research file into a draft."
- "Checker: compares the draft against the sources and flags anything unsupported."
When a bot has one job, it's obvious where the work goes next. It also makes each bot easier to test and fix, because you know exactly what it is supposed to do.
3. Write hand-off notes
Each bot finishes by leaving a short note for the next one: what it did, where the output is, what is still open and what the next step should be. Something like:
Done: collected sources for the article, saved to research file. Open: two sources disagree on a date. Next: Writer drafts the article and flags the date for the Checker.
It feels small, but it removes most of the questions I used to answer by hand.
4. Keep a person at the checkpoints that matter
Handing work between bots doesn't mean removing people. At Novatise, a person still approves every article before it is published. The goal is to stop being the copy-and-paste layer, not to stop being responsible for the result.
What changes
The main change is where my time goes: I spend it deciding what the bots should do and checking what matters, instead of moving text between windows.
If you're building your own team of bots
Ask yourself one question: if you went on holiday for a week, could your bots pass work to each other without you? If the answer is no, start with a shared file and clear job titles. That alone removes most of the copying.
This is the same approach we use when we build business process automation for clients at Novatise: map who does what, give each step one job, and connect the steps so people only step in where their judgement is needed. We can usually show a working prototype within 24 to 48 hours.
Stuck being the glue between your own tools and AI agents? Talk to us at novatise.com about connecting them into one working system.