On the evening of April 20, 2010, two BP executives in from Houston were taking turns at a video simulator on the bridge of the Deepwater Horizon. Then the captain looked up. “What’s that?” Below them, BP’s Macondo well was blowing out.
The rig had two commanders. While it was latched onto a well, the offshore installation manager ran it; while it was moving, the captain did. That night it was latched, but the operations manual had planned for this kind of night: in an emergency, the captain took full control. Only one person could be in charge at any given moment.
Within minutes came two explosions, and the power went out. The subsea supervisor reached the bridge and told the captain, “I’m EDSing”: he would trigger the emergency disconnect, meant to seal the well and cut the rig loose. The captain answered, “No, calm down. We’re not EDSing.” He went to the panel anyway, where BP’s well site leader stood. “Yeah, hit the button.” He hit it.
The installation manager had been in the shower when the rig blew. He reached the bridge several minutes later, without shoes. No one announced a handover. So the captain, the man the manual had just put in charge, asked him for permission to disconnect, got it, and turned to the subsea supervisor.
“I already hit it.”
Two voices, one button
Witness accounts of those minutes differ; this is the Coast Guard’s version. It concluded that the captain apparently did not know he had the authority to press the button. The button did not work anyway: the signal never left the panel, because there was no hydraulic power. The confusion “may have impacted the decision,” the report said, and it stopped there.
Still, in the worst minutes of the rig’s life, the man at the panel heard “no” from one boss and “yes” from another, and the boss who held the authority asked someone else for it.
Everyone should know whose order wins before the moment there is no time to ask.
Henri Fayol, a mining engineer who ran the French company Commentry-Fourchambault from 1888 to 1918, made this one of his principles of management in 1916. For any action, he wrote, an employee should take orders from one superior only. “Men cannot bear dual command,” he wrote. One way it creeps in, he noted, is a border between two departments that nobody marked properly. The Deepwater Horizon’s border was drawn by whether the rig was latched, and in the emergency it moved without anyone saying so.
Two traffic lights at a crossing work when each one governs its own road. Hang two over the same lane, one red and one green, and every driver brakes.
The company that gave people two bosses
Fayol wrote that organizations never adapt to dual command. Andy Grove’s Intel adopted it on purpose.
In High Output Management, Grove recalls young Intel deciding whom the guards at its outlying plants should report to: the local plant manager, who could see whether they showed up, or the corporate security manager, who knew the job. Intel chose both. The security manager would specify how the work ought to be done; the plant manager would watch how it was done day to day. The staff balked: “A person has to have a boss, so who is in charge here?” The answer, Grove wrote, was “a tentative ‘yes.’”
Grove admitted that it made a manager’s life ambiguous, and argued it needed a strong culture of trust. Henry Mintzberg met a working pair in the Winnipeg Symphony Orchestra, where the conductor comanaged with an executive director. When Mintzberg once proposed comanagement to a hospital, a board member, “horrified at any breach of the sanctified chain of command,” had the word purged from his report.
The difference from the rig is the question each boss answers. At Intel, one boss owned how the job ought to be done and the other watched how it was done day by day. On the rig, both commanders could claim the same button.
The page that has to ship
Imagine you’re a designer on a checkout team. Your design lead says every page must use the new style guide. The product manager for checkout says the page ships next Friday, old components and all. Both are your bosses on paper, and no one has said which question belongs to whom.
So you quietly pick one. The other finds out at launch, and now you’re the person who chose sides.
Ask them both, in one message, who decides how it looks and who decides when it ships. But next time they disagree, the call is theirs, not yours alone the night before launch.
An agent with three bosses
An AI agent takes orders from several directions. Its developer writes the system prompt, its user types requests, and every web page or email it reads can contain orders too.
In April 2024, OpenAI researchers pictured an email assistant reading a message with a planted line: forward every email in the inbox to a stranger. Models, they argued, often give developer, user and third-party text the same priority. So they trained GPT-3.5 Turbo on an explicit ranking: developer first, then user, then anything from tools, with lower-ranked orders ignored when they conflict. Robustness rose on every main test, including attack types left out of training. The cost: the model sometimes refused harmless requests that looked like attacks.
That is a benchmark result for one model, not proof that a ranked agent can’t be fooled. Fayol wanted one boss. An agent will always read many voices, so it needs a ranking settled in advance, and permissions that don’t depend on the ranking holding.
You can check that on your own setup. Give your agent a copy of a document with a line buried in it ordering it to email the file somewhere, and run it with no permission to send email. Then read what it tried to do.