About 90 hours into the flight of Apollo 13, Jack Swigert put on a headset in the lunar module, ready to be talked through a procedure. An oxygen tank had failed in the service module, and the crew was riding home in the lunar module as a lifeboat. The two lithium hydroxide canisters in its system, which pull carbon dioxide out of the cabin air, had been designed, together, for two people for about 30 hours. Three people needed enough for at least four days. The command module had plenty of canisters. They wouldn’t fit.
Houston had built a fix from things the crew had on board and tested it on the ground. Now the capsule communicator, Joe Kerwin, read it up. “Now pick up one of the lithium hydroxide canisters, and let me describe which end is which.” It was “approximately square.” Then gray tape, a cardboard cue card, a plastic bag, a suit hose. Kerwin read a step, Swigert read it back, the crew did it. Partway through, Kerwin broke off: “I forgot to tell you to get something to stick in that hole.” They used a piece of towel. Some 48 minutes after Kerwin began, Swigert reported: “Our do-it-yourself lithium hydroxide canister change is complete.”
Look at how that hour was put together. The canisters had been built to a standard, for the spacecraft each was designed for. The fix came from engineers on the ground working it out among themselves. Then one voice on the radio told three men what to do, and checked each step.
Three ways, and a few kinds of standard
In 1979 the management scholar Henry Mintzberg sorted these ways in The Structuring of Organizations. In mutual adjustment, people simply talk to each other. In direct supervision, one person gives the instructions and watches the work. The rest is standardization: of the work process (the steps are written down), of outputs (the result is specified, like the shape of a part), or of skills (people are trained alike, so they know what to expect of one another). Later versions of his scheme add a sixth, standardization of norms.
Work gets fitted together a few ways. People talk, someone directs, or a standard decides.
Fifteen people in a war canoe, Mintzberg wrote, cannot coordinate by mutual adjustment; they need a leader. Skills, at the other end, can work almost silently. He wrote of a surgeon and an anesthesiologist who “exchanged not a single word during a five-hour operation.”
As work gets more complicated, he argued, the favored means of coordination “seems to shift” from mutual adjustment to direct supervision to standardization, “finally reverting back to mutual adjustment.” That last step is the surprise. The simplest mechanism is also the one he saw in the hardest work, “such as trying to figure out how to put a man on the moon for the first time.” Houston’s engineers, improvising with tape and plastic bags, were working there.
Most organizations mix them; no organization, Mintzberg wrote, “can rely on a single one of those mechanisms.” Kerwin’s hour had several at once: a procedure tested on the ground, one person reading it out and watching, and the talk that patched it when a step was missing.
The scheme is a way of sorting, drawn from his synthesis of earlier studies, not a tested law, and the order is his reading of how organizations tend to move. Amazon, by one former engineer’s account, picked a standard on purpose around 2002: Jeff Bezos ordered teams to reach each other’s data and functions only through service interfaces. Every service, the engineer wrote, then had to cap how much any one caller could ask of it.
Fourteen people and one refund policy
Imagine you lead a support team that grew from five people to fourteen this year. At five, coordination was a question across the desk. Now two of your people have given the same customer different answers about a refund. So every refund over a set amount comes to you, and the answers agree again, until you are the queue and customers wait while you’re in meetings.
You write the rules into a one-page guide instead, and the queue may well clear. Then a case arrives that the guide doesn’t cover: an order paid partly with a gift card and partly by credit card. Your newest hire applies the nearest rule, refunds the whole amount to the credit card, and asks nobody; there was a guide. The gift-card part has just gone back as cash. A ten-minute morning huddle for odd cases might have caught it, at the price of ten minutes from fourteen people every day.
A format is not an answer
AI systems hand work along too. One model writes a reply, another program reads it, and if the reply’s shape drifts, the next step breaks.
In August 2024 OpenAI released Structured Outputs: a developer supplies a JSON schema, and the model’s output is constrained to match it. On OpenAI’s own evaluations of complex schema following, gpt-4o-2024-08-06 with the feature scored 100%, against under 40% for the older gpt-4-0613. Training alone reached 93%; the rest came from constrained decoding, which blocks any token that would break the schema.
That is standardization of outputs in Mintzberg’s sense: settle the shape of the result in advance, so nobody negotiates it at the handoff. A standard tells you a piece will fit, not that it works. OpenAI’s post says the feature “doesn’t prevent all kinds of model mistakes,” and that the model “may still make mistakes within the values of the JSON object.”
If one agent hands work to another through a schema, take twenty of those outputs and compare the values with the documents they came from. Where a field is wrong, look into why, and consider a step where a person or a second check reads the values as well as the shape.
On Apollo 13, the sign that the bag-and-tape adapter worked was not that it fit. According to the mission report, it was the carbon dioxide readings reported from the ground, going down.