Every n8n flow is an agent we froze by hand.

The whole company runs on n8n flows. Look closely at one: a trigger fires, then each node does one thing — read Airtable, call an API, post to Slack — and hands its result to the next, until the last node produces an output. That is exactly what an AI agent is: a set of tools plus instructions. The only real difference is who decides the order. In an n8n flow a human drew every arrow in advance. In a real agent the model decides — every run. This page shows, in plain language, how a flow maps onto a prompt + tools, and why our flows aren’t true agents yet.

One Maestro, many agents, each a bundle of workflows.

Here is the whole target in one picture. Maestro sits on top — it holds the global governance, the rules and the settings (it’s where the Governor gate lives). Under it, one agent per domain: Gabrielle for sales, Robin for R3 ops, Frank for finance, Jacob for junk. Under each agent live its workflows — but a workflow here is no longer an n8n canvas: it’s a skill, a prompt that carries the same instructions the flow used to encode. Every agent reaches the outside world through one shared tool set, and humans steer the whole thing from a single UI layer.

What a real agent is

An agent is two things: a prompt (the goal plus the rules) and a set of tools (each with a name and a little schema of inputs). You hand the model the goal; it picks a tool, reads the result, picks the next, and stops when the goal is met. Nobody draws the arrows — the model derives them, live, from the goal and from what it sees on each step.

The driver-exception flow, both ways

Here is a real Robin flow — a driver messages that something’s wrong at a site — written first as the n8n we run today, then as the agent it could be. Same tools. Same goal. Different driver.

# system prompt
You are Robin's driver-exception handler.
A driver reported a problem at a site.
Goal: figure out which order(s) it
affects, classify the exception, and file
one review task for Joan.
If you can't match an order, file a
"no-order" task instead. Never file twice.

A flow is an agent’s transcript, frozen

We drew the reasoning once, by hand, and saved the drawing. A real agent re-derives that reasoning on every run — so it handles the cases we never drew. Same tools. Same goal. The model holds the pen instead of us.

The agents, and what sits on top

A companion to the question Jess asked — every AI agent in Agent Settings, organized by department, each with the lead agent that sits on top and the subagents below it, plus the Maestro runtime that holds it all together. Click any agent in the map for detail; dashed cards are proposed (build gaps).

Where an agent actually lives and runs

An agent is a definition an admin writes in a console — name, instructions, model, tools, skills, secrets, a memory setting — that gets published and then run on a serverless runtime. That machinery is the Agent Control Plane (ACP): a console, a runtime, and the storage that holds both the definitions and the record of every run. Everything is done from the UI — no deploy.