In some professions, currently, value is shifting towards the ability to orchestrate complex systems rather than produce them oneself. (Illustration generated by Gemini/Nano Banana Pro)

In some professions, currently, value is shifting towards the ability to orchestrate complex systems rather than produce them oneself. (Illustration generated by Gemini/Nano Banana Pro)

As companies rush towards “agentic” organisations, a new figure is emerging: the AI agent orchestrator, a human in charge of supervising, correcting and taking responsibility for the errors of autonomous systems. A revolution in the workplace that questions the balance of power between humans and machines.

The rise of so-called “agentic” organisations, described in detail by McKinsey in a report this autumn, marks a profound shift in the way companies think about work. This model is based on swarms of autonomous AI agents capable of executing entire sections of a process: analysing documents, preparing budgets, reviewing contracts, auditing data, launching campaigns or orchestrating technical tasks. In this scheme, the human being is no longer the operator but the supervisor: he or she “regulates the agents”, checks for deviations, arbitrates contradictions and bears ultimate responsibility. McKinsey describes three human archetypes: “M-shaped supervisors”, generalists responsible for steering automated workflows; specialists who catch exceptions; and workers in direct contact with the public, now supported by proactive systems. This typology reflects an operational reality: value is shifting towards the ability to orchestrate complex systems rather than produce themselves.

The report points out that the power of agents could grow rapidly: from the ability to perform around two hours of reliable work in 2025, systems could reach several days of autonomy by 2027. This scenario remains uncertain and depends on as yet unverified technical hypotheses, but it is already fuelling the restructuring of organisations. In this context, the figure of the “agent orchestrator” takes on strategic importance: a few humans would be capable of supervising dozens, and then potentially hundreds, of specialised agents. McKinsey mentions the example of a human team of two to five people managing an “agent factory” of 50 to 100 agents covering an entire process. However, this projection is based on feedback from early-stage companies.

This vision is shared, at least in its general intuition, by several publicly accessible academic studies. A 2024 MIT study on the behaviour of autonomous agents in work processes highlights the risk of a shift towards intensive coordination work, where humans spend more time managing emergent agent behaviour than performing tasks of direct value. The more functional autonomy systems gain, the more demanding human supervision becomes.

Papers deposited on arXiv in 2025 - including “The Era of Agentic Organization: Learning to Organize with Language Models” and “Future of Work with AI Agents” - also explore the idea of hybrid teams made up of autonomous models and supervising humans. These studies analyse prototypes of ‘agentic workflows’ in controlled environments, where agents perform successive micro-tasks under the supervision of a human operator. They suggest that this type of configuration increases productivity in certain simple cases, but that it makes the cognitive load more complex as soon as agent-agent interactions become numerous. For the moment, there is nothing to suggest that these conclusions can be generalised to large-scale industry.

The “moral crumple zone”, a place of humanity... and responsibility

The Boston Consulting Group published a report in 2025 on the management of autonomous systems in the enterprise (“Managing the Machines”) which stresses the need for new human skills centred on arbitration, operational ethics and the management of risks induced by agents capable of taking limited initiatives. The report also refers to shifted responsibility: in automated systems, humans remain accountable even when they have not generated the system’s actions themselves. This phenomenon has already been theorised in the research of Madeleine Clare Elish with the concept of the “moral crumple zone”, where humans absorb the fault of the system without really controlling all its parameters. This analogy is useful for understanding the pressure that could be placed on agent orchestrators in future organisations.

Throughout these sources, the same dynamic emerges: the AI agent orchestrator becomes the backbone of systems where models execute, iterate and talk to each other, while the human remains ultimately responsible for consistency, meaning and compliance. This reversal of logic - the human is no longer the doer, but the one who is accountable for the consequences of what the machine does - can be read as a natural evolution of advanced automation. But it can also be analysed as a new form of dependency in which a few workers pilot swarms of autonomous systems that, in effect, impose their own pace, limits and blind spots.

The “slave driver” metaphor in the title obviously refers neither to history nor to any kind of moral comparison, but to an inversion of the apparent power relationship: humans appear to be in charge, but it is they who absorb the mental load, the legal responsibility and the risk of operational drift. The agents, on the other hand, work without fatigue, without breaks, without time limits. The human supervisor thus becomes a manager of continuous flows, assessed on his ability to maintain the stability of a system that operates faster than he does.

There remains one major unknown: will this emerging profession be attractive, sustainable and truly value-creating? None of the available sources currently provides robust empirical data on the well-being, cognitive load or real organisational impact of these functions. What is certain, however, is that the shift towards agentic organisations will not be possible without massive investment in the human skills of supervision, arbitration and governance. The future of work will not just be automated: it will be orchestrated, and this orchestration will have a human cost that companies are only just beginning to contemplate.