Twenty airline crews, two pilots each, came to a NASA simulator to fly a twin-jet on a trip modeled on their own airline’s operation. Half arrived within two or three hours of finishing a real three-day, high-density, short-haul duty cycle. The other half had usually been off duty for three days, a few of them for only two.
The flight was built to go wrong. The weather was poor and getting worse. At the destination, the crosswind rose above the legal limit for the approach, the runway stayed hidden in cloud, and every crew ended up abandoning the landing. During that missed approach one of the jet’s hydraulic systems failed. The gear and flaps had to be lowered by hand, the brakes were weaker, and once the gear was down it couldn’t be raised again. The only reasonable place left to land was a short, wet runway, 5,800 feet long, sloping downhill next to mountains.
The crews coming off the trip were, not surprisingly, significantly more fatigued. Yet their performance was, in general, significantly better. An expert observer rated them better on a number of dimensions relevant to flight safety. The NASA researchers called the pattern “somewhat counter-intuitive.”
The variable nobody had set
Then they noticed something about who was sitting in the cockpit. Crews at the end of a trip had usually flown the whole trip together. Crews at the start often had not. Crews whose two pilots had flown together on their last duty cycle averaged 5.0 errors in the simulation. Crews whose pilots hadn’t averaged 11.67.
The re-sorting came after the fact, with twenty crews, and nobody had been assigned to fly together or apart. But in 1994 the National Transportation Safety Board added evidence from real accidents. Of the major U.S. airline accidents from 1978 to 1990 that involved the flight crew, it could tell for 15 whether the pilots were new to each other. In 11 of them, 73 percent, the accident came on the captain and first officer’s first day flying together. The Board said that lent support to the NASA study.
You can’t make a team great. You can set up the conditions that make it more likely.
That is how the Harvard psychologist J. Richard Hackman saw it. He told Harvard Business Review in 2009 that “even the best leader on the planet can’t make a team do well. All anyone can do is increase the likelihood that a team will be great by putting into place five conditions.” In his 2002 book Leading Teams, he named them: a real team, a compelling direction, an enabling structure, a supportive organizational context, and expert coaching. By his 2011 book Collaborative Intelligence the list had six, adding the right people.
A gardener can’t make a plant grow. She can choose the soil, the light and the water, and the plant does the growing.
Who is on the team?
The first condition sounds too obvious to need saying. A real team, in Hackman’s sense, has a shared task, clear boundaries, and members who stay together long enough to learn how to work with each other. Airline cockpits often fail the third test, by design. Airlines get the most out of planes and pilots, Hackman explained, by scheduling each one separately. An airline operations researcher once estimated for him how long it would take before two pilots who had flown one trip together could expect to fly together again: 5.6 years.
Boundaries fail more quietly. In a study of more than 120 senior leadership teams, Hackman and his colleagues found that almost every team thought its membership was clear. Asked to describe their team, fewer than 10 percent of members agreed about who was on it.
How strong is the evidence for the whole set? Much of it comes from Hackman, his colleague Ruth Wageman and their co-authors, in field studies that measure the conditions and outcomes rather than assign teams to them at random. Wageman writes that the six conditions, as measured by their Team Diagnostic Survey, “together account for 80% of the variance in how well a team performs.” That is the developers’ own figure for their own instrument.
A direction a model can read
Some conditions carry over to AI agents, some don’t. A model doesn’t get tired and doesn’t need to be inspired. A “compelling” direction, for a model, mostly means a specified one.
In a study first posted in 2025, Chenyang Yang and colleagues at Carnegie Mellon looked at what happens to requirements a prompt leaves out, across three representative tasks. Models often met them anyway: in 41.1 percent of cases, an unstated requirement was satisfied more than 98 percent of the time. But those guesses were fragile. Across model updates, unstated requirements were almost twice as likely as stated ones to lose more than 20 percent in accuracy. One small update to GPT-4o, between two versions a few months apart, made it on average 48 percent less likely to “produce skimmable outputs.”
Writing everything down was no cure either. The authors found that simply specifying every requirement did not consistently help, because models follow only so many instructions and requirements can conflict.
What you can try: list the requirements your prompt relies on but never states. Turn the important ones into checks you run on real outputs. When the model version changes, run the checks again before trusting the old results.
The group that met every week
Imagine you’re asked to rescue a cross-department project that keeps slipping. Nine people from three departments, each on it part-time, meet every Monday. Your first instinct is an offsite to build some spirit.
Instead you ask each of the nine, privately, to write down who is on the team and what the team, not each department, is supposed to deliver. You get back four different lists of names and three different deliverables.
The offsite can wait. The next Monday, the meeting starts with a list everyone can see.