An agent is not a chatbot with tools. It is a system that takes actions with consequences — which makes the only useful question how far up the autonomy ladder it is standing, and whether the controls for that rung exist yet.
Agentic systems are usually specified by what they can do. They should be specified by what happens when they are wrong.
Every action inside that boundary needs a limit somebody wrote, a log somebody reads, and a stop somebody has actually pulled.
A wrong draft costs a minute. A wrong payment costs a relationship. The same architecture is not appropriate for both.
Reversibility is a design property, not an incident-response hope. If an action cannot be reversed, the gate in front of it has to be real.
An agent is not a chatbot with tools. It is a system that takes actions with consequences, and the only useful question is which rung it is on. Choose a rung.
Five levels of autonomy. Each one is a different system with different consequences, and each one has a control that has to exist before you stand on it.
We will not build an agent on a rung whose controls do not yet exist. That is not caution for its own sake — it is that the rung below is almost always where the value actually was.
Four stages, every engagement. Hover a stage to see what happens in it.
Three moves. Two of them are ours, and the one in the middle is deliberately somebody else’s.
Design, integration and operation of the system, with a named human accountable at every gate that matters. This is delivery work and we do not pretend otherwise.
The assurance opinion on anything we built is not written by the team that built it, and where you need it to carry weight externally, not by Orvix at all. A firm that audits its own delivery is offering you a marketing document.
We fix what the independent review says is wrong, and the fix is re-examined by the same reviewer rather than signed off by us.
Three engagements inside this pillar. Start with the question you can name, or take the whole estate at once.
Systems that act, with a named person accountable at every gate.
The limits, logs and kill switches that make autonomy reversible.
The inventory, lineage and quality the models are actually standing on.
Thirty minutes. Describe the action you would least like to explain afterwards, and we will start there.
Book a 30-minute scoping callORVIX · INDEPENDENT AI & TECHNOLOGY ASSURANCE · WE DISCLOSE EVERY COMMERCIAL RELATIONSHIP ON THE PAGE FOR THE SERVICE IT BELONGS TO. WHERE LICENSING IS REQUIRED, DELIVERY IS PERFORMED BY NAMED PARTNERS UNDER THEIR OWN LICENCE.