The game takes 18 minutes. Four people get 20 sticks of dry spaghetti, a yard of tape, a yard of string and one marshmallow. The tallest structure that stands on its own wins. The marshmallow has to be on top.
Tom Wujec, who builds tools for design teams, ran it in about 70 workshops around the world. Most teams, he said in a TED talk in 2010, go about it the same sensible way. They discuss the task, plan, sketch and lay out spaghetti. They spend most of their time building one tall structure. Near the end, someone places the marshmallow on top, everyone steps back to admire it, and the weight of the marshmallow brings the whole thing down.
Among the worst at the game, Wujec found, were recent graduates of business school. Among the best were recent graduates of kindergarten.
The children didn’t know more about building. They went about it in a different order.
Marshmallow first
The children started with the marshmallow. They built a small tower with it on top, and when that held, they built another, taller one, always keeping the marshmallow on top. “With each version,” Wujec said, “kids get instant feedback about what works and what doesn’t work.”
The business students, he said, were trained “to find the single right plan,” and then carry it out. The plan got tested once, in the last minute, when there was no time left to fix it.
The best teams of all were architects and engineers, who know that triangles hold their shape. Knowing the material still counts. But the order counted enough for kindergarteners to beat people with business degrees.
Wujec saw the same thing from the other side when he raised the stakes. He offered one group of design students $10,000 worth of software for the winning tower. A one-inch structure would have won. Not one team had a structure left standing. When the same students played again, building and testing as they went, they went from the very worst to among the very best.
Write down the guess
In 1939 the statistician Walter Shewhart argued that specification, production and inspection “must go in a circle instead of in a straight line.” W. Edwards Deming took the idea to Japan in 1950, and Japanese executives, by one account, recast it as plan, do, check, act, the version most people know.
Deming never warmed to that name. “Check,” he warned, can mean “to hold back.” In a 1990 letter he wrote: “Be sure to call it PDSA, not the corruption PDCA.” The S stands for study. Plan a change. Try it, preferably on a small scale. Study what happened: what did we learn, what went wrong? Then act: adopt it, drop it, or go round again.
The difference he cared about is in the study step. Ronald Moen, who worked with Deming for years, and two colleagues made it explicit in 1991: the plan has to include a prediction, written down before you start, and the study compares what happened with it. A good check can do that too. But without a guess on paper, it is easy to look at the result and decide it was roughly what you expected.
One lunch
Imagine a café manager tired of the lunch queue going out the door. A full redesign of the counter would take months. So one Tuesday she tries a single change. A second person walks the queue with a notepad and takes orders before people reach the till. On a sticky note by the register she writes her guess: waits under eight minutes. And she asks a colleague to jot down, for every customer, the time they joined the queue and the time their order was taken.
At half past one she reads the sheet. Waits are down, though not to eight. And the kitchen, used to orders arriving one at a time, got them in bursts. Two salads went out wrong.
So the guess was half right, and the slow spot had moved to the kitchen. She keeps the notepad and decides that tomorrow the orders will go back to the kitchen in threes. The whole experiment cost one lunch.
The loop inside an agent
An AI agent can go round this loop again and again within a single task. Anthropic’s guide to building agents describes them as models “using tools based on environmental feedback in a loop,” and calls it crucial for an agent to get “ground truth” from the environment at each step, such as the output of a tool or the result of running code.
Study starts when that real result is set against what the agent expected. An agent asked only to judge its own work is like a manager checking the plan against the plan. The result of running the code is the sheet at half past one.
Anthropic’s guide to its coding agent, Claude Code, puts the cost plainly. Without a check the agent can run, “you become the verification loop: every mistake waits for you to notice it.” The check doesn’t have to be a test suite. The guide lists a build’s exit code, a linter, a screenshot compared against a design. What matters is that the signal doesn’t come from the agent’s own opinion.
The kindergarteners never needed a judge. Every few minutes the marshmallow sat on top of a new tower, and the tower either stood or it didn’t.