It's not the variety. It's the simultaneity.

One morning this spring, the daily brief produced by my orchestrator agent contained a line I hadn't asked for. During the night, one of the AI agents it coordinates had seen fit to widen its scope of analysis to a segment of accounts we had deliberately set aside. The initiative was argued, the logic held, the work was clean. It was also contrary to a decision I had made three weeks earlier, for reasons the agent could not know, because they appeared in no document it could read.

I corrected, reframed, wrote the missing rule. And as I did, I recognized the move. It was the one I had made a hundred times with brilliant, overeager junior salespeople. Take the work back, explain the context the other person could not have guessed, put the limit in writing so the situation would not repeat itself. The entity was new. The move was not.

The first paper of this triptych laid out the chronology: four regimes in six years, a craft that changed in nature. This one goes down into the day-to-day. What do you actually do when you manage, all at once, humans, humans who manage other humans, humans who manage AI agents, autonomous AI agents, and AI agents that manage other agents?

The answer holds in two observations that seem to contradict each other. Almost everything has changed. And the moves that let you hold on are the oldest in the craft.

A day in five registers

First, we need to describe what is genuinely unprecedented about this daily life, because the keynotes never talk about it with any precision.

My morning starts with a brief nobody wrote. An orchestrator agent assembled it overnight from the work of several specialized AI agents. One sorted the inbound signals, another updated the files of active accounts, a third prepared the day's priorities according to rules I have written and amended dozens of times. That brief is waiting for me when I wake up. It condenses work that would have taken me two days to produce alone in 2019.

Then the day shifts into another register. I brief a senior salesperson on a live negotiation, and there everything is human: the unsaid, the energy, the trust to be dosed. An hour later, I go over an account plan with the manager who prepared it, and I am managing a manager, which is yet another language. In the afternoon, I adjust a control rule that a team lead applies to the AI agents in his chain, because an agent output brushed against a compliance limit the week before. And at the end of the day, I reread the orchestration logs of an agent system to understand why a processing chain made an unexpected decision.

Five registers in a single day. Five languages, five rhythms, five criteria of evaluation. The senior salesperson expects recognition and a frame. The manager expects autonomy and perspective. The AI agent expects nothing, but demands written rules of a precision I had never had to produce for a human. And the agent system asks for something else again: an architect's reading, almost an urban planner's, where you no longer look at an entity but at flows.

*Managing in 2026 is not doing a new job. It is doing five jobs before lunch.*

What has changed: the material

Then we have to say honestly what the ambient discourse does not say. AI agents hallucinate. Not marginally, not only in competitors' failed demos. In production systems, regularly, with an aplomb that makes the error more dangerous than a human's. A human who doubts shows it. An AI agent that is wrong argues.

I have seen a qualification agent produce an impeccable analysis on data it had partially invented, because a source had stopped responding and it had filled the gap without flagging it. The report was fluid, coherent, convincing. It was wrong on part of its content. Other times, it is an invented number in the middle of accurate data, or a cited source that does not exist. No human collaborator produces this type of error. A human who does not have the data says so, apologizes, works around it. The agent completes. It is a new category of error, one you do not detect by reading but by checking the sources.

Agents also take unsolicited initiatives, like the one in my opening. They come in varied forms: a widened scope, a reprioritization decided on their own authority, content that overflows the instruction. It is the exact flip side of what makes them valuable. You give them autonomy because it multiplies their value, and that autonomy produces, by construction, decisions you have not validated. A strictly obedient AI agent is an expensive script. An autonomous AI agent is an entity that will, from time to time, step out of the frame. There is no perfect setting between the two. There is a permanent arbitration, which shifts with the tasks, the stakes, the maturity of the system.

And all of this happens at a speed that rules out exhaustive supervision. The orchestrator that briefs me every morning coordinated, overnight, a volume of work I cannot fully recheck. Nobody can. So the question is no longer how to control everything. It has become: what to control, at what frequency, and with what guardrails for everything else.

*A manager who endures his AI agents is not an augmented manager. He is an equipped spectator.*

What has not changed: the moves

Faced with this material, two reactions are settling in around me. Those who look for a revolutionary method, an "AI agent management" to be learned as a brand-new discipline, with its gurus and its express certifications. And those who quietly abdicate, who let the systems run and tell themselves the technology knows what it is doing.

Both are wrong in the same way. They believe the newness of the material demands newness of the move. Yet the moves that hold have existed for forty years in the most classical management literature. What has changed is what they cost when you neglect them.

Ritual, first. My morning brief is at a fixed time. My review of agent outputs is weekly, scheduled, non-negotiable, exactly as my Monday pipeline review was in 2012. An AI agent system without a supervision ritual drifts like a sales team without a pipe review. Not out of malice, through the accumulation of small deviations nobody catches. I learned at my own expense that a week without reviewing the orchestration rules is a week in which three micro-drifts settle in and reinforce one another. The ritual is not bureaucracy. It is the metronome that makes drift visible before it costs.

The framing of autonomy, next. Andy Grove set down in High Output Management a principle I hold to be the most useful ever written on delegation: the level of supervision should depend not on a collaborator's general competence, but on his maturity on the precise task at hand. Task-relevant maturity. An excellent performer on a new task calls for close control. The same one on a mastered task calls for being left alone. I applied that principle for fifteen years with my teams. One salesperson, excellent at closing, remained closely supervised on his forecast; another, solid in prospecting, had free rein on his territory but not on pricing. Trust was never granted wholesale. It was granted task by task.

Grove was writing for managers of human engineers, in 1983. Yet his principle describes, word for word, the proper management of an AI agent in 2026. My agents have radically uneven maturity across tasks. On document synthesis, I hardly check anymore; months of reliable outputs have established trust. On edge cases, sensitive segments, anything touching an external commitment, I check everything, systematically. The only difference with 1983 is that an AI agent's task-relevant maturity evolves in months, not years, and can regress from one update to the next.

*Grove's principle has not aged. It has accelerated.*

Mastery of time, finally. Peter Drucker opened The Effective Executive with a chapter every manager should reread once a year: Know Thy Time. Effective executives, he wrote, do not start with their tasks; they start with their time, because time is the only resource that can be neither bought, nor stored, nor recovered.

There is a scene in the documentary Inside Bill's Brain that long served me as a reminder. Bill Gates's assistant says of him that he is on time to the minute, at every meeting, without exception. And she adds that time is the only resource this man, who can buy everything, cannot buy. His day is twenty-four hours long, like ours, and nothing will ever change that. I have often told that scene to my teams. One of the richest men on the planet treats his calendar with more rigor than any of us, because he understood before we did that it is the one thing that cannot be bought back.

I was saying it in my own words well before the first AI agent: a technically excellent salesperson who cannot manage his time will end up behind an average salesperson with sound work hygiene. Competence does not compensate for structure. That principle has simply changed scale. AI agents free up considerable time, and that is precisely what makes them dangerous. A manager without structure fills that freed time with noise: compulsive checking, one more tool, one more dashboard. I have seen managers gain fifteen hours a week thanks to agents and end up more overwhelmed than before, because nothing in their organization decided what those fifteen hours were for. Drucker would say they gained time without knowing it. The gain evaporates when no structure is there to receive it.

*Time remains the key variable of success. AI agents have not changed its nature. They have raised its price.*

The fundamental that becomes central

One discipline remains, and it is the one that now separates the managers who hold it together from those who wear themselves out.

In the stable regime before 2020, the will to keep learning was a quality. It separated the good from the very good, but you could make a career without it, carried by what you had learned once. In the current regime, it has changed status. It is no longer a quality. It is the condition for practicing the craft.

The material changes every year. The AI agents I manage today do not resemble those of eighteen months ago, and next year's will make some of my current settings obsolete. Whoever has not made his own relearning an organized discipline, with blocked time, framed experiments, documented errors, is not falling behind. He is leaving the craft, slowly, without noticing, while continuing to do well what used to matter.

And this is where the two observations of this paper meet. The fundamental moves hold. Grove holds, Drucker holds, the pipeline review holds. But the fundamental that governed all the others has changed address. It was not control. It was not even time. It is the capacity to relearn faster than the material changes.

*The fundamentals of management were not replaced by AI. They were sorted by it.*

What that sorting does to the craft as a whole, what it demands of those who still want to practice it in ten years, and what new discipline it sketches out, is the subject of the third and final paper.

Article 2 · Sales management triptych · July 2026