Etiennator is the AI agent system that runs our client work, and it is laid out as a building. Each client has its own floor, where agents read that client's Slack messages, emails and call notes, turn them into tasks and do the work.
Every result goes past a head agent before it reaches me. What passes lands on one Decide list, with a push each morning, where I accept, send back or drop each result.
The short version
- Etiennator is our in-house AI agent system: each client has an isolated floor of agents that reads its Slack, mail and call notes and turns them into tasks.
- Agents do the work in a sandbox, and a head agent checks every result before it reaches my Decide list.
- Six months ago the same kind of work ran on fixed n8n workflows, and every yes was me clicking in n8n.
- It does not yet judge whether a task was worth doing, and nothing client-facing goes out on its own.
One floor, from a client message to my click
The client writes
Slack, mail and call notes, for that client only
The floor makes tasks
what it reads becomes a list of tasks
Departments do the work
agents work in a sandbox
The head agent checks
every result is reviewed before I see it
I decide
Accept, Send back or Drop, from the Decide list
Read it in order: the first four steps run without me. In the Mac app the same system is drawn as a 3D building, floor by floor.
What changed: from n8n clicks to a Decide list
Client work now starts on its own, and what reaches me is a decision.
I have built 200+ n8n workflows, and each one is a fixed chain of steps that runs when its trigger fires. A workflow does exactly what it was drawn to do, so anything new needed me to notice it and draw another workflow.
Take a client request in Slack. Six months ago it waited until I read it. Today the floor turns it into a task, a department does the work, the head agent checks it, and what reaches me in the morning push is a finished result.
Six months ago
- A client request waited until I read it
- Automation meant fixed n8n workflows that ran when a trigger fired
- Every yes was me, clicking in n8n
Today
- The client's floor turns the request into a task on its own
- Departments do the work and a head agent checks it
- What reaches me has already been checked once
Since 9 October, an agent is free to take its next task while its result waits on the Decide list. Before that, a finished task kept its agent waiting on a decision.
Who does what inside the AI agent system
A floor has departments that do the work, a librarian for the client's records, a head agent that checks, and me.
Floors are isolated: each client's messages, records and work stay on that client's floor. Each floor keeps its own records on its client, and since 29 September all 5 floors have had a librarian agent for them.
The head agent follows check 4 in how to know if AI automation is working: whoever does the work should not be the one who approves it. What passes its review goes on my Decide list, next to the plans agents want my OK on before they start.
Two standing rules of mine apply to the agents' work. Nothing client-facing is sent automatically: the agents do the work and never send the message. And on my own data they add things but never change or delete what was there before.
What comes next: fewer clicks
The next step is the click itself: today every result waits for me, and an agent should be able to earn its way past that.
We have written the rule for it. An agent earns the right to close its kind of work alone once enough of its results have matched my decisions, and until it does, everything it produces waits for me.
It is the same path we set for one client's messaging work: a person approves everything, then spot-checks, then the work runs alone. The sign I am waiting for is a head agent that asks whether a result makes sense, as well as whether it works, before it reaches my list.
What it does not do yet
It does not judge whether a task was worth doing, and the head agent can still pass work that is wrong, so I still check what reaches me. Whether a task was the right thing to do is my call, the gap I wrote about in AI now works before I ask. It also runs on one approver, so it moves at the speed I read my Decide list: two floors already stop suggesting new work while 10 or more finished results wait for me.
Key takeaways
- An AI agent system can be organised like a company: one isolated floor per client, departments that do the work, and a head agent that checks it.
- Agents that read a client's own channels can start the work without waiting for a person to read the message first.
- One list of decisions, each with three buttons, lets one person stay in charge of many agents without reading everything they do.
- Hard rules sit above the agents: nothing client-facing is sent automatically, and nothing in my existing data is changed or deleted.
Common questions
What is Etiennator?
Ellipse Automation's in-house AI agent system for client work. Each client has an isolated floor of agents that turns its messages into tasks and does the work, a head agent checks every result, and Etienne Chanut approves from a Decide list.
Do the AI agents send anything to clients on their own?
No. Nothing client-facing is sent automatically: the agents do the work and never send the message, the head agent checks it, and Etienne accepts, sends back or drops each result.
What does the head agent do?
It reviews every result a floor produces before Etienne sees it, so what reaches him has already been checked by something other than the agent that did the work.