Agent Factory and Expert in the Loop: what an agentic enterprise looks like

Dr. Martin Hofmann was Group CIO at Volkswagen for nine years, then spent several years in Silicon Valley and was most recently CTO at the electric truck start-up Volta Trucks. He is writing a book about the agentic enterprise.

The way in is an image that sticks: at a Swiss mobile phone retailer, agents with their own name and their own staff number sit at the virtual table in some meetings.

The four stages

Hofmann pins the difference between chatbot and agent on a sequence that works well as an aid to classification.

First data analytics: evaluating what is there. Then machine learning: recognising patterns and predicting. Then the LLM chatbot: answering questions. And finally agentic AI: accessing data independently via interfaces such as MCP, carrying out tasks and learning from its own behaviour.

The decisive break with classic IT thinking lies not in the technology, but in the way the assignment is given.

The Agent Factory

Hofmann's practical concept is called the Agent Factory: employees from IT, from the business side and from human resources design their own digital colleagues in a protected practice area instead of waiting for consultancies or systems integrators.

The approach solves a problem that every wave of automation has had: whoever knows the process cannot build it, and whoever can build it does not know the process. Until now that gap was bridged by requirements documents, with a familiar result.

A protected area is the precondition. Without it, the same thing arises as with Excel macros: distributed, unchecked automation with nobody responsible.

Control, concretely

Control does not stay lip service with Hofmann, and the two means can be named.

The Agentic Journal logs every prompt and every data source. That is the replacement for the traceability that arose in the classic process through approval stages. Without this log it cannot be said afterwards why an agent did something.

The kill switch takes effect when the deviation from the defined result becomes too great. That presupposes that somebody has laid down beforehand what too great a deviation means. That too is a business task, not a technical one.

Hofmann's correction of terms is notable. He considers “Human in the Loop” the wrong frame and coins Expert in the Loop instead: the human assesses the quality of the digital employees rather than monitoring them.

The difference is more than cosmetic wording. Monitoring means looking at every single operation, and that does not scale. Quality assessment means judging samples and patterns, and that is exactly the activity specialists are trained for. It is at the same time the activity of a manager towards employees, which follows the image of the digital colleague through to the end.

What that means for roles

Hofmann's book “The Agentic Enterprise” with a twelve module framework addresses CIOs, CDOs and CHROs. That the head of human resources appears in it is the most instructive part.

When agents get staff numbers, sit in meetings and are answerable for results, they are no longer an IT topic. Questions then arise about induction, appraisal, development and decommissioning, and those are HR processes.

Conclusion

This episode delivers the most usable ordering model for the question of how a company works with agents.

Three points can be checked straight away. Do you specify results or process steps? If process steps, you are using an agent like a script and giving away the capability.

Is there a log showing which data an agent has seen and which instruction it was given? If not, you have surrendered the control that previously sat inside the process.

And who assesses the quality? If the answer is “everybody keeps an eye on it”, nobody does.

Hofmann's advice to the next generation fits the rest: not social sciences but mathematics, because logic and adaptability remain the core competencies in a world that delivers a new model every five days.