My method — supervising AI agents
In one sentence: one day with you, one day with Alice — you create, you deploy, but never without a safety net. I analyse, I process, I guide you step by step. A full architecture or a plain language model plus a control tower: it depends on the feel, the mistakes, the fixes, the guardrails. And cockpits, so you can always see what your agents are doing, inside which boundary, with which rights.
The rhythm: one day with you, one day with Alice
Nothing runs unattended. The rhythm is simple, and it always comes back around:
- one day of work with you — the agent, the model;
- one day of work with Alice — the local agent, who learns on my machine and remembers outside the model.
During those days you can start to create and to deploy, but safely: never without a net, never without a trail.
What a day is for
A day is first of all about understanding your trade — how you reason, where you get stuck, where you shine. Then, in order:
1. I analyse the data you produce,
2. I process it,
3. I guide you through the assessment, step by step, with Alice.
This is not monitoring for the sake of monitoring. It tells us where to put the next guardrail, and where to give you a longer leash.
What can come out of it
Everything is on the table, and it gets decided as we go:
- an architecture of your own — circuits, gates, dedicated tools;
- or something lighter: a language model plus the control tower.
Nothing is fixed in advance. The system gets built while walking.
What makes the difference
It all depends on the person and the relationship, not on a specification document:
- the feel, the exchange, the conversation;
- the mistakes and the fixes;
- the automations and the repetitions;
- the guardrails, the tools, your circuits.
Visibility, always
We have plenty of visual tools to help: cockpits, and Packet-Tracer-style visualisation — it looks good, and above all it is useful.
We do everything so that, at every level, you keep:
- visibility on what your agents are doing,
- control over which boundary they work in and which rights they hold.
An agent you cannot watch working is an agent you can neither trust nor correct. Here, you watch.