Intent, agent performance, human check: leading when everyone has 109 agents

Leadership has so far worked like conducting. The corporate strategy is the score, the talents sit in the orchestra, and the manager's job is to make something harmonious out of it in which everyone can develop. That image carried for a long time.

It is tipping right now. Dr René Deist, a guest for the third time, points to a Gartner forecast according to which every employee will soon have 109 agents at their disposal. If that holds, nobody is conducting any more. Then every individual in the orchestra has to formulate a strategy themselves, that is to say compose.

The model behind the division of labour

Deist's starting point has been unchanged since their first joint episode and carries the whole argument.

“AI wants nothing. AI has no intent. AI also has no chance of assessing a risk within itself, because AI does not feel it. But we humans do.”

From this follows a three-way split that he now calls intent, agent performance and human check. Played through using procurement as an example, it looks like this: a human determines which raw material is to be negotiated in what volume with which suppliers at what time. That is the intent, and it is a strategic statement, not a work instruction.

The agents then work. They write to suppliers, evaluate replies, spot gaps, follow up, conduct the correspondence. At the end stands the large table with all the offers.

Then a human looks at it again and notices that the screw factory named cannot possibly deliver the promised quantity. Precisely this step cannot be delegated, because it rests on world knowledge and on a sense of risk that no model brings along.

What becomes of the job profiles

From the model Deist derives an educational objective, and it is more concrete than the usual calls for further training. Anyone working in procurement today will develop in one of two directions over the coming years: towards master of intent or towards master of check and control.

In between sits the question of who actually builds the agents. Deist's answer shifts the responsibility: the specialists themselves have to learn to break their work steps down so that skills come out of them. IT supplies the framework, that is the execution environment, the checking mechanics and the connection to the systems.

Note what this means for the organisational structure. If business units describe their own automations, it takes conventions, storage locations and approval routes for them. Otherwise the same shadow IT arises as back in the days of Excel macros, only with considerably greater reach.

The paradox: more leadership, not less

The widespread expectation is that automation reduces the need for leadership. Deist counters with a calculation that makes immediate sense: when two departments work together, in future it is not two people sitting at the table but 218 agents. Each individual leads their own fleet, and these fleets have to be coordinated with one another.

That automation simultaneously means doing more tasks with fewer people, he states openly. The two together do not make a comfortable picture, but an honest one.

How quickly such a thing gets out of hand when a stopping criterion is missing is shown by an example from the episode: a loop running over the weekend opened 4,800 Electron instances, and the agent involved concluded at the end that it could not repair that either. That is the practical side of agent performance without a human check.

Prompt thinking and its limit

The book's second thesis is called prompt thinking. Deist illustrates it with a question he hears frequently: how do you intend to ensure quality when AI runs processes autonomously, when five attempts at having a presentation built already failed.

His objection: “Make me a PowerPoint with a strategy for something” delivers a random result. Four pages of prompt with role, experience background, data source, structure and desired look deliver a controllable one. The difference lies not in the model but in the specification.

There is, however, a restriction that the episode makes explicitly: prompt thinking works where you have mastered your craft. Anyone who has not built good slides before will not get them from a machine either, because they cannot write the four pages of specification. The technology lifts existing competence, it does not replace it.

AI literacy as a leadership task

Deist's advice to managers is level-headed: AI literacy belongs on the leadership agenda, no more and no less than the use of the pocket calculator did in its day. Yes, people have been worse at mental arithmetic since then. It was right nonetheless.

For him a particular understanding of leadership belongs to this.

“Leadership is a service to people.”

In practice that means: hire people who are cleverer than you, question yourself constantly, correct the course when it is wrong, and give your staff room to experiment. In an environment where everyone is learning at the same time, it is no weakness if knowledge flows from the bottom upwards.

Deist sees the risks somewhere other than in a digital two-class society, which he disputes for the western world. He considers two other points more serious: engaging with the technology too late, and dependence on those who master it and later set the prices. A petition of around 300 signatories, among them 15 Nobel laureates, accordingly calls on governments to lead structural change rather than stop it.

Conclusion

The practical core of the episode is a question of responsibility, not of technology. As long as it is clear who formulates the intent and who owns the check, the middle part scales at will. If one of the two roles is missing, the problem scales with it.

For implementation that means: for every automated process, set down in writing who formulates the assignment, how success is measured and who looks at it at the end. Those three lines matter more than the choice of framework. And define the stopping criterion before the first loop runs over a weekend.