On the morning of Wednesday, January 22, 2003, the space shuttle Columbia was on its seventh day in orbit. Six days earlier, a piece of foam had broken off the External Tank during launch and struck the left wing. Nobody on the ground had a picture sharp enough to say what it had done.
Wayne Hale, a Shuttle Program manager at Kennedy Space Center, called a Department of Defense representative at Kennedy and asked the military to start planning to photograph Columbia in orbit. Within an hour the question had reached US Strategic Command, where a planner began identifying assets that could do it.
The call had gone around the chain. It was placed without authorization from Linda Ham, chair of the Mission Management Team, and to a representative who was not the designated liaison for such requests. Ham called members of her team and asked whether there was a “requirement” for imagery. None of them had requested it or could identify one.
At 8:30 a.m., NASA’s Defense Department liaison called Strategic Command and cancelled. The request had been made and rescinded within 90 minutes. A follow-up e-mail asked that future requests be “confirmed through the proper channels.”
A plank across the ladder
Nearly ninety years earlier, a French mining engineer drew this problem. In Administration industrielle et générale, published in 1916, Henri Fayol pictured an organization as a double ladder with the chief at the top. Suppose F, on one side, needs to settle something with P, on the other. By the line of authority, the question climbs from F up to the chief, descends rung by rung to P, and then makes the same trip back.
It is much simpler, Fayol wrote, to go straight from F to P across a footbridge. He called it the passerelle; the 1949 English translation calls it a “gang plank.” Its use is “simple, swift, sure.” Two employees can settle in one sitting, in a few hours, a question that by the chain would go through twenty transmissions, inconvenience many people, involve masses of paper and lose weeks or months.
Peers across branches can deal directly, if both bosses allow it and hear what was agreed.
Fayol attached exactly those conditions. The chain is respected if F’s boss and P’s boss have authorized them to deal directly, and the matter is regularized if both report back at once on what they agreed. “From the instant that agreement ceases or there is no approval from the superiors,” the translation reads, “direct contact comes to an end, and the scalar chain is straightway resumed.”
Hale’s call had no approval from the manager who owned the mission. When Ham found out, the contact ended and the chain resumed. Fayol’s condition, that direct contact stops once a superior doesn’t approve, had worked as written.
Who builds the plank
The rule worked. The question didn’t get through. Engineers on the Debris Assessment Team, the people analyzing the strike, had wanted pictures too; Ham never asked them, the accident board wrote, whether the request was theirs. The team’s NASA co-chair drafted an e-mail he never sent: “Remember the NASA safety posters everywhere around stating, ‘If it’s not safe, say so’? Yes, it’s that serious.” He didn’t want to jump the chain of command.
Fayol had an opinion about such silence. Avoiding the plank is usually blamed on fear of responsibility, he wrote, but he thought the cause was insufficient executive capacity in the people in charge. If the top man insisted that his deputies use the plank and made their subordinates use it too, the habit and courage of responsibility would follow. “It is an error to depart needlessly from the line of authority,” he wrote, “but it is an even greater one to keep to it when detriment to the business ensues.” Then he added a conditional exception. When an employee must choose between the two and cannot ask a superior, “he should be courageous enough and feel free enough to adopt the line dictated by the general interest.” For that, he wrote, there must be precedent, and superiors must have set the example, “for example must always come from above.”
The channels had a reason. The liaison’s e-mail said the chain exists so that requests don’t slip through and stir up worry over problems not yet vetted. Side channels also carry more than warnings: the board found that a tile expert used his personal network to shape the managers’ view and, in its words, “snuff out dissent.” A plank carries whatever walks across it.
The fix nobody upstairs heard about
Imagine you work in finance and notice that a sales team’s pricing sheet applies a discount twice. You message the analyst who owns the sheet, and the two of you fix it in twenty minutes. Simple, swift, sure.
Neither of you tells your manager. Next quarter the sales director, who had approved the double discount for three large accounts, finds their invoices higher than promised and has the old sheet restored. Nobody upstairs knew the plank had been crossed, so nobody could say why it had been.
Agents that message each other
In Claude Code’s agent teams, an experimental feature that is off by default, one session acts as lead and spawns teammates. Anthropic’s documentation contrasts this with subagents, which report results back to the main agent. Teammates, it says, “communicate directly with each other,” and any teammate can message any other by name.
That is Fayol’s footbridge. The lead decides who is on the team; reporting back is the condition to check. The documentation describes a finished teammate’s final answer reaching the lead in an idle notification; it doesn’t say that messages between teammates are copied to the lead. Each agent’s mailbox is a JSON file under the team’s folder in your home directory, and the docs say the team config directory is removed when the session ends.
So ask for the report. Tell each teammate, in its spawn prompt, to send the lead one line whenever it agrees something with another teammate. Before the session ends, open the mailbox files and compare what the teammates told each other with what the lead told you.
For Columbia there were no mailbox files to open. The accident board pieced the side calls together months later, from logs, e-mails and personal notes, after the shuttle broke apart on re-entry on February 1.