In the last week before the Mars Climate Orbiter reached Mars, the number kept dropping. The plan was for the orbiter’s closest pass to be 226 kilometers above the planet. As tracking data came in, the prediction fell, and an hour before arrival it was as low as 110.
On September 23, 1999, the main engine fired, and four minutes later the orbiter passed behind Mars, 49 seconds earlier than predicted. Its signal never came back.
The navigators at NASA’s Jet Propulsion Laboratory had been uneasy since April. That was when they finally got usable versions of a file from Lockheed Martin Astronautics in Denver, which built and operated the orbiter. The file listed the size of every small thruster firing, the routine nudges that kept the craft’s flywheels from spinning up and pushed it slightly off course each time. Within about a week, the data suggested the nudges were being underestimated. Through spring and summer, people at the working level worried. The investigation board later found their concerns “only informally reported.”
There was a cross-check sitting in the data the whole time. The spacecraft calculated each nudge itself, in the right units, and radioed the result to Earth. The navigators didn’t learn about that until long after launch. They had assumed this craft was much like an earlier one.
Two pictures of one spacecraft
Teams coordinate better when everyone carries the same picture of the task and each other.
At the Mars Climate Orbiter, there were at least two pictures. The navigators’ was of Mars Global Surveyor, the earlier spacecraft. As the board put it, they expected much of the new craft’s hardware and software to be similar. It wasn’t. The new craft’s solar array was asymmetrical, so sunlight built up more momentum that the thrusters had to dump, and a daily flip meant to cancel that out had been dropped from the operations plan. Nobody told the navigators. The thrusters fired 10 to 14 times more often than they expected.
Then there was the file itself. A ground program in Denver wrote the size of each firing into it. The interface specification said newton-seconds. The program wrote pound-force seconds. Every nudge entered the navigation model 4.45 times too small, and the error sat exactly where one team’s work met the other’s.
The board found the navigators had not communicated their trajectory concerns effectively to the spacecraft team, and that the spacecraft team had not understood them. The operations navigators had come on board shortly before launch and had taken part neither in testing the ground software nor in the design reviews.
Psychologists had a name for what was missing. In 1990, Janis Cannon-Bowers and Eduardo Salas extended the idea of a mental model to teams, to explain expert teams that coordinated smoothly without needing to talk it through. In a 1993 chapter with Sharolyn Converse, as Lim and Klein later summarized it, they proposed that a team is most likely to be effective when members share four: of the equipment, of the task, of how the team interacts, and of one another’s knowledge and skills. The navigators’ model of the equipment was, in part, of a different spacecraft.
Three agents and a bomb
Teams of AI agents have the same gap, because an agent knows only what is in front of it.
In a 2023 study, Huao Li and colleagues put three GPT-4 agents, Alpha, Bravo and Charlie, into a text game. They had to find and defuse color-coded bombs in a set of rooms, each agent holding different wire cutters and seeing only its own room. Anything one learned, the others knew only if it said so.
The agents did coordinate. But they also acted on a state of the game that wasn’t true: searching for a bomb already defused, claiming a bomb’s sequence without inspecting it. Once, Alpha told the team it was moving to a room it couldn’t actually reach.
The researchers’ fix was to prompt each agent to keep a written note of what it believed about the task, updated after every observation. GPT-4 teams defused every bomb either way. With the notes, they did it in 12.3 rounds on average instead of 28.3. That was one small game, three trials per condition, with a 2023 model.
You can borrow the move. When several agents share a job, have each write its working picture into a file you can read: units, assumptions, what it believes is done. Read the files side by side. Where two disagree, check one figure against something independent, the way the orbiter’s own numbers could have been checked against the tracking data.
How shared is shared
A 2010 meta-analysis by Leslie DeChurch and Jessica Mesmer-Magnus pooled 23 studies and found shared mental models positively related to team performance, however they were measured. In a field study of 71 Singapore Army combat teams, Beng-Chong Lim and Katherine Klein found that both the similarity of members’ models and their accuracy, judged against experts’ models, predicted the teams’ assessment-center scores three weeks later. These are associations, not experiments.
Agreement isn’t the same as accuracy. A whole team can share the wrong picture and agree on it comfortably.
The board’s advice for the next Mars mission, the Mars Polar Lander, was practical. Bring an attitude-control specialist onto the navigation team. Hold a 15-minute stand-up meeting every morning, so everyone could hear what was happening.
Two numbers for one quarter
Imagine you’re preparing the quarterly review. Sales reports 1,200 new customers. Finance reports 950. Both teams are sure of their number, and both are right: sales counts a signed contract, finance counts the first paid invoice.
You put both definitions on the slide, and then you take a single customer, one who signed in March and paid in May, and walk it through both systems from end to end.
On the slide, the two numbers now sit side by side, each with one line underneath saying what was counted.