On a Friday night in August 2016, in Rio de Janeiro, the American 4x100-meter relay team was carrying the flag around the Olympic track. They had just crossed the line third. Their second leg was Justin Gatlin, the 2004 Olympic 100-meter champion.
One place ahead of them was Japan. None of Japan’s four sprinters had a personal best under ten seconds for 100 meters. Japan had the lead going into the last baton change, and only Usain Bolt, anchoring Jamaica, pulled away. Japan’s time, 37.60, was an Asian record.
The Americans were still on the track when they found out. Officials ruled that the leadoff runner, Mike Rodgers, had passed the baton to Gatlin outside the first exchange zone. The medal was gone, lost in the few meters where one runner hands off to the next.
“We had great sticks in practice, great everything, and something so simple,” said Tyson Gay, who ran the third leg. By the Associated Press’s count, it was the ninth time since 1995 that the U.S. men had been disqualified or failed to get the baton around at the Olympics or world championships.
Talent minus the handoffs
A team delivers what its members could do, minus what gets lost working together.
The social psychologist Ivan Steiner wrote it as an equation in his 1972 book Group Process and Productivity. Actual productivity equals potential productivity minus “losses due to faulty process.” Potential is the most the group could produce with what its members bring. For a rope pull, as later researchers put it, that’s the number of people times what an average member pulls alone. The process is everything that happens between the members, and in a relay a lot of it happens in the exchange zones.
Japan’s federation went after the losses. Since 2001 its sprinters had used an underhand baton pass. “The foremost reason for using the underhand pass is that it is least prone to errors,” Shunji Karube, the Japanese federation’s sprint director, told World Athletics. Relay practice started in March, with exchanges drilled in every running order, months before the four names were settled. Between the heat and the final, Karube moved the outgoing runners’ start marks by seven centimeters.
Steiner also went back to an older case. In the 1880s a French engineer named Ringelmann had measured workers pulling on a rope, and groups pulled less than the sum of their members. Steiner named one source of the loss, poor coordination, and showed that in Ringelmann’s data the shortfall grew in step with the number of coordination links between pullers. Later experiments showed some of the loss on a rope is people easing off, a second leak described on the Ringelmann Effect page. Both come off the same top line.
The equation has limits. It has only a minus sign. The team researcher J. Richard Hackman argued that the right conditions can also open up “synergistic process gains.” And potential is easy to estimate for a rope or a relay, and hard for most work. Nobody can add up what a product team could do at its best. Karube himself didn’t treat process as the whole answer. “We need sub-10 seconds sprinters,” he said after Rio.
Sixteen agents, one bug
In February 2026, Nicholas Carlini, a researcher at Anthropic, described setting 16 Claude agents loose on one job: write a C compiler, in Rust, that could build the Linux kernel. Each agent worked in its own container and claimed tasks by writing a lock file into a shared repository, so two of them wouldn’t take the same one.
While the work was hundreds of separate failing tests, the agents spread out. Then they reached the kernel, and in Carlini’s words it was “one giant task.” Every agent hit the same bug, fixed it, and overwrote the others’ changes. “Having 16 agents running didn’t help because each was stuck solving the same task.”
Steiner had a word for that kind of task: unitary, one that can’t be split into separate subtasks. Carlini’s fix, in effect, made it divisible. He used GCC, an established compiler, as a known-good reference: compile most of the kernel with GCC and only a random set of files with the new compiler. If the kernel worked, the problem wasn’t in that set. If it broke, the set could be narrowed down. Each agent could then chase a different bug in different files. It didn’t remove the hard part entirely. Some files failed only in pairs, and finding those took more debugging.
Adding agents raises the potential on paper. Whether any of it arrives depends on whether the work splits. On a copy of your project, run the same job with one agent and with several, a few times each, and read the history: how often two agents edited the same file, or fixed the same failure. A lot of overlap is a lead worth checking against the task before you add more.
The fifth analyst
Imagine you run a team of four analysts who build the quarterly forecast. It came in late again, and you’ve asked for a fifth person.
Before the request goes up, you ask each of them to log one week. The pattern is plain. Each analyst spends about three days on their own region, and then the team spends two more reconciling: different definitions of “active customer,” a currency column one of them converted and the others didn’t, a formula overwritten in the merge.
Four people make six pairs who have to agree on things. Five make ten. The new hire would bring a fifth spreadsheet to the same merge.
So for one quarter you try something smaller: one shared template with the definitions written at the top, and one person who owns the merge. The request for a fifth analyst stays in your drafts folder, waiting on the next week of logs.