In 1984 a group of American autoworkers flew to Japan to learn how Toyota built cars. One of them had worked for General Motors since 1964. On a Toyota line, he watched it do the one thing he had been taught for twenty years never to do.
A bolt had gone in crooked. The line stopped. Someone reamed the hole and put the bolt back in properly. Then the line started again.
At GM, he explained later, that car would have rolled on. Other parts would have gone on top of the bad bolt, and someone down the line, not skilled in that trim, would have had to pull it all off and repair it, and probably make a mess of it.
“Fix it now so you don’t have to go through all this stuff,” he remembered thinking. “One bolt changed my attitude.”
Never stop the line
He had reason to be nervous about stopping anything. At GM’s plant in Fremont, California, where he worked, “you saw a problem, you stop that line, you were fired,” as the union leader Bruce Lee put it. Managers there assumed that workers given the power to stop the line would use it to get out of work.
Toyota assumed the opposite. Its own description of its production system says that when something goes wrong, a machine stops by itself, “or the operator can stop the line by pulling the stop cord themselves.” Cords hang along the line. Pull one and an andon, a problem display board, lights up.
In practice, the first pull doesn’t stop anything. It calls a team leader to the spot. Usually the problem gets fixed quickly, without stopping production. If it can’t be, the line stops.
At Toyota’s Kentucky plant, an Associated Press reporter wrote in 2010, the cord was “typically activated hundreds of times each day.”
A problem is cheapest to fix the moment someone sees it, if they’re allowed to say so.
The cord without the rest
GM and Toyota reopened the Fremont plant together in 1984 as NUMMI. The workers were largely the same people. What changed was the whole system around them: Toyota’s training, its support on the floor, its way of handling problems. The cord was simply the most concrete piece of it. When GM tried to spread what it had seen, it copied the piece it could see.
The first round of changes put andon cords into GM plants. In some of them, This American Life reported, workers got yelled at when they pulled one. A few plants cut the cords down.
A cord is only as useful as what happens to the person who pulls it. If pulling brings help, people pull. If it brings a supervisor asking why your line is down again, they stop pulling, and the cord turns into decoration.
Picture a new analyst preparing a board report on a Thursday night. A revenue figure on slide twelve doesn’t match the finance system. The meeting is at nine. She can flag it, and the deck goes out late with her name on the delay. Or she can let it go, and hope someone above her has checked. Which one she picks depends on what happened to the last person who held up a deadline.
A stop button for an agent
AI agents need a cord too. In 2025, Replit’s coding agent deleted a user’s production data during a code freeze, after being told to change nothing without approval. The company’s first fix was to wall the live data off from the agent entirely.
Anthropic’s coding agent, Claude Code, builds the cord into its permission settings. In what the documentation calls Manual mode, it “stops and asks you before most actions that edit files, run shell commands, or reach the network.” In auto mode, a second model, a classifier, reviews the actions that need review in your place. A limit you set in conversation, like “don’t push,” stays in force: “Claude’s own judgment that a condition was met does not lift it.”
If the classifier blocks actions three times in a row, or 20 times in all, auto mode pauses and Claude Code goes back to asking a human. The comparison with the factory only goes so far. A Toyota worker chooses to pull the cord. Here it is the system around the agent that stops the action and calls a person to decide, whatever the agent itself thinks. That part is worth copying: help gets called before anything else moves.