By 1888 the Commentry-Fourchambault company, a group of coal mines and ironworks in central France, had paid no dividend for three years. The coal from its Commentry mine was “far from paying its cost of extraction,” according to the man who ran that mine. By his later account, the company was “almost resigned to closing and abandoning its factories and suspending operations in the mines.”
That man was Henri Fayol, who had joined the company at 19, straight from mining school, and stayed with it. He knew coal, not iron. The board wanted its Fourchambault works closed. As Ellen O’Connor’s 2015 review of a history of the company reports it, Fayol first withdrew his name rather than close that works. The board appealed to his devotion. He took the job on one condition: Fourchambault would not close until his own investigation showed it had to.
Twenty years later, in 1908, he stood before fellow mining engineers in Saint-Étienne and described a company that had recovered. It hadn’t done it with new mines or new people, he insisted. “It was with the same mines, the same factories, the same employees that the Company went into decline before 1888 and with which it has recovered since that date.”
The decline and the revival, he said, “were due solely to the administrative procedures used.”
Five verbs
In 1916, at 75, he published Administration industrielle et générale in a mining society’s journal (it came out as a book the next year) and defined the job in one line: “Administrer, c’est prévoir, organiser, commander, coordonner et contrôler.” To manage is to plan, organize, command, coordinate and control.
Managing is a craft of its own, beside the work itself; Fayol named five parts of it.
The French verbs don’t map neatly onto the English ones. Prévoir was to scan the future and draw up a plan of action; the 1949 English translation needed two words for it, “forecast and plan.” Commander, in his definition, was to keep the staff working. Coordonner was to bind together and harmonize every act and effort. Contrôler was to see that everything happens according to the rules set and the orders given; Jean-Louis Peaucelle and Cameron Guthrie’s 2015 history of his career, Ellen O’Connor notes in her review of it, translates it as “verifying.”
He described planning in “a large mining and metallurgical firm” with about ten thousand employees. Two months after each year closed, a general report covered it, with forecasts for the next. A ten-year forecast was redone every five years, so a course was always traced at least five years ahead.
In 1908 he had told his audience of engineers that a leader who administers well but is a mediocre technician is generally more useful than a brilliant technician who administers badly. A position, he said, is “a sort of lever, which can multiply his personal influence by two, by ten, by 100 or even by 1,000.”
A relay team of four fast sprinters can still lose on a dropped baton. The running is the technical work. Who runs which leg, the practiced handoff and the check at the line are the manager’s part.
His own account
Fayol’s story about 1888 is his own, and it leaves things out. As Commentry’s coal ran down, the company bought mines at Brassac in 1891 and Decazeville in 1892, according to a Montluçon local history society. Its plant at Imphy developed specialized steels for new markets. O’Connor points out that he wrote the theory decades later, looking back: “This is not a scientific process.”
Henry Mintzberg went further in Harvard Business Review in 1975. Fayol’s words, he wrote, “tell us little about what managers actually do. At best, they indicate some vague objectives managers have when they work.” The executives Mintzberg watched spent their days in short, broken pieces.
Fayol listed what the work is for; Mintzberg logged the hours. The list is easier to test where every step leaves a written record.
Same model, different management
Teams of AI agents leave that record. In 2025 a team mostly from UC Berkeley, with Mert Cemri among its lead authors, studied why systems in which several language-model agents split a task fail. They collected 1,642 execution traces from seven open-source multi-agent frameworks, mostly running GPT-4 and Claude models on coding, math and general tasks, and sorted the failures into three groups: system design (vague specifications, roles not followed), misalignment between agents, and task verification.
Laid beside Fayol’s list, the groups fall roughly on planning, coordination and control. In one trace a phone agent never told its supervisor agent that the username had to be a phone number; the supervisor never asked, and the login failed again and again. In another, a chess program passed superficial checks such as compiling, but nobody validated it against the rules of chess, and it had bugs when it ran. The paper’s first insight echoes Fayol’s 1908 talk: a well-designed system “can result in performance gain when using the same underlying model.”
Most traces were labeled by a language model calibrated against human experts, and these are research frameworks, not every agent product. The authors’ fixes to one framework helped, but, they wrote, did “not constitute substantial improvements.”
Open the log of a run where your agents got something wrong, find the step that approved the result, and write down what it actually checked. If it only confirmed the code ran, run the output against the task’s own rules and see what it would have caught.
The bug the new lead took
Imagine you’ve just been made lead of the six-person team where you were the strongest engineer. Two weeks before a release, the nastiest bug arrives, and you take it yourself, because you’ll fix it fastest. Three late nights, and it’s fixed.
On release day you find that two of your engineers each built the export feature, differently, because each thought it was theirs. And the database migration was only ever run on test data.
By midnight you’re rolling the release back. Your fix goes with it.