On October 30, 1935, the US Army held a flight competition at Wright Field in Dayton, Ohio, to pick its next long-range bomber. Nobody expected much of a contest. Boeing’s Model 299 had already beaten its rivals in early tests. It could carry five times the bombs the Army had asked for and fly almost twice as far as earlier bombers. A newspaperman had nicknamed it the “flying fortress.”
The plane took off smoothly, climbed to about 300 feet, stalled, turned on one wing and crashed in flames. Two of the five crew died, including the pilot, Major Ployer P. Hill.
The investigation found no mechanical failure and called it pilot error. Hill had forgotten to release a new lock on the elevator and rudder controls. The plane had four engines, retractable landing gear, new wing flaps, electric trim tabs and propellers whose pitch was set with hydraulic controls. Hill had a lot to attend to. A newspaper called it “too much airplane for one man to fly.” The Army gave the contract to Douglas, and Boeing nearly went bankrupt.
The obvious fix was more training. But Hill had been the Air Corps’ chief of flight testing. It was hard to imagine a pilot with more experience. So a group of test pilots tried something much simpler.
They wrote a checklist.
1.8 million miles
With the checklist, by the surgeon and writer Atul Gawande’s account, pilots flew the Model 299 a total of 1.8 million miles without one accident. The Army went on to order almost 13,000 of the planes. It called them the B-17.
Hill hadn’t lacked knowledge. He had too much to hold in his head at once. Even a step you know by heart can get skipped on a busy day.
Memory is a net with holes, and the more you carry, the bigger the holes get. A checklist is a second net, held under the first.
Gawande cites a study from Israeli intensive care units that shows how big the holes can be. The average patient needed 178 separate actions a day. Staff got 99% of them right. That still worked out to about two errors per patient, every day.
Could a sheet of paper do the same for doctors? Many physicians doubted that a page of little boxes would help much.
Five steps every doctor knew
In 2001 Peter Pronovost, a critical-care doctor at Johns Hopkins, picked a single problem: infections from central lines, the tubes run into large veins. He wrote five steps on a sheet of paper. Wash your hands. Clean the skin with antiseptic. Cover the patient with sterile drapes. Wear a mask, hat, gown and gloves. Put a sterile dressing on the site.
These were basics that had been taught for years. Making a checklist for them seemed almost silly. He asked nurses to watch for a month anyway.
In more than a third of patients, at least one step was skipped.
The next month he changed one rule. The hospital gave nurses the authority to stop a doctor who skipped a step. A year later, the ten-day line infection rate in his unit had gone from 11% to zero.
Plenty of doctors were unconvinced, and some were offended at the idea that they needed a checklist at all. Johns Hopkins had money, staff and Pronovost walking the halls. So the test moved to Michigan, to more than 100 intensive care units, including hospitals where staff were stretched thin. There the checklist came with something else: senior executives who visited each unit and found, among other things, that the antiseptic soap was stocked in fewer than a third of them. The study recorded the median infection rate falling from 2.7 per 1,000 catheter days to zero within three months, and a drop of up to 66% over a year and a half.
Operating rooms told a similar story. In a study that ran from late 2007 into 2008, eight hospitals from Seattle to Tanzania started using a 19-item surgical checklist from the World Health Organization. Among adults having non-cardiac surgery, the study recorded deaths in hospital within 30 days falling from 1.5% to 0.8%, and complications from 11% to 7%.
The empty box
Picture a Friday afternoon in an accounts office. The most careful person on the team has thirty invoices to send before the weekend, a phone that keeps ringing and a manager asking about a late payment. Invoice twenty-two goes out without the client’s purchase order number. On Monday the client’s system rejects it, and the payment slips a month. She has sent hundreds of invoices. She knows the number has to be there. On that afternoon, she was carrying too much.
A long list of everything tends to become a form people tick without reading. The lists that worked in these stories were short and limited to the steps that actually get skipped. Someone was allowed to say “stop” when a box was empty. And someone made sure the soap and the drapes were on the shelf. Take those away and a checklist is a piece of paper on a clipboard.
A checklist with teeth
An AI agent seems like the last thing that needs a checklist. It doesn’t get tired or distracted.
It does have its own version of a busy day. Anthropic’s guide to its coding agent, Claude Code, warns that as a long session fills up, the model may start “forgetting” earlier instructions. It also says, bluntly: “Claude stops when the work looks done.”
The guide offers a checklist in two strengths. You can write the steps into the instructions, which works like Pronovost’s sheet of paper. Or you can add a script that runs the tests and won’t let the agent stop until they pass. The guide calls written instructions “advisory.” It calls these automatic gates, which it names hooks, “deterministic.”
The first is a reminder. The second is the nurse, rebuilt in software, standing at the door and saying “you’re not finished.”