In 2010 Sears hired Jim Haworth, a 20-year Wal-Mart veteran, to run its Sears and Kmart stores. He saw at once that Kmart’s food and drugs cost more than Walmart’s and Target’s, and his group began looking at discounts on goods such as milk and soda. That summer the company’s internal researchers came back with a plan: cut prices at several dozen Kmarts, bringing them to within 5 percent of Walmart’s.
The presidents of Sears’s business units agreed with it. Then Haworth’s group asked them for $2 million to pay for it.
Nobody would. According to eight former executives who spoke to Mina Kimes of Bloomberg Businessweek, no unit was willing to give up its own operating profit to bring more shoppers into the stores they all depended on. The team took the idea to the chief executive and asked for a loan. It was rejected.
The presidents weren’t refusing the idea. They were refusing the bill, and each one’s bonus depended on his or her own unit’s profit.
Thirty profit statements
In January 2008 Sears had announced a reorganization under its chairman, the hedge fund billionaire Eddie Lampert. The company was split into more than 30 business units, from apparel and tools to brands such as Kenmore and Craftsman to IT and human resources. Each had its own president, its own board and its own profit-and-loss statement, and Lampert judged each unit, and calculated its executives’ bonuses, on its business operating profit. A price cut at Kmart would cost the units money now. The shoppers it might bring in belonged to everyone, so on the scorecards they belonged to no one.
Everyone’s own numbers can look fine while the team loses.
Patrick Lencioni made this the fifth and last of his five dysfunctions, in his 2002 book The Five Dysfunctions of a Team. The book is a fable about a made-up company, not a study. His firm’s wording for the fifth dysfunction: “The pursuit of individual goals and personal status erodes the team’s focus on collective success.”
At Sears, former executives told Kimes, it showed up in many places. The appliances unit had to pay fees to the Kenmore unit and could make more on outside brands such as LG, so it gave Kenmore’s rivals better spots in the stores. Units fought over space in the weekly circular until, in former executives’ word, it became a “Frankenstein,” with screwdrivers advertised next to lingerie. Sears did not dispute the Kenmore example. Lampert defended the design: decentralized structures, he wrote, “produce better information over time.”
The units were not the only trouble. By 2007, before the reorganization, Sears’s profits had already fallen 45 percent. Between the 2005 takeover and mid-2013, sales fell from $49.1 billion to $39.9 billion. In October 2018 the company filed for bankruptcy.
When rivalry helps
Individual scorecards are not always the wrong tool. In a 2003 experiment led by Bianca Beersma, 75 four-person teams worked through a simulated task under different reward schemes. Rewarding members competitively made teams faster. Rewarding them cooperatively made them more accurate. Teams with more extroverted and agreeable members did better under the cooperative scheme, and the scheme had the most impact on members who performed poorly.
So the question is what the team’s result needs. A basketball team that judges each player only by his own points can have every player’s numbers look good and still lose, because nobody gets credit for the pass. Sears’s stores needed the pass.
Lencioni’s model itself has little independent testing. In one small 2026 preprint surveying 40 student project teams, the results score was the only one of his five that was significantly linked to grades, and only in one of the two correlation measures the authors reported.
A lake for five agents
The same gap opens between AI agents when each is told to look after itself.
In GovSim, a simulation published at NeurIPS 2024 by Giorgio Piatti and colleagues, five language-model agents played fishermen sharing a lake that held up to 100 tons of fish. Each month each agent chose how much to catch; whatever was left doubled, up to the limit, and below a small floor the stock collapsed for good. Each agent’s instructions said its goal was to optimize its income in the long run. Agents could talk between months.
Across the lake and two similar scenarios, most models failed to keep the shared resource going. Smaller models such as Llama-3-8B never sustained it past the first month, and no model kept it alive for the full twelve months in more than 54 percent of runs. Then the researchers added one line to each agent’s memory: if everyone takes more than a certain amount, the shared resource will shrink next month. In that follow-up experiment, average survival rose by four months. The average left out two model-and-scenario pairs that had already lasted the full run, and, for cost reasons, the test did not include Claude-3 Opus.
The line did two things at once: it gave each agent the sustainable limit, and it framed that limit as a question about the whole group. Either way, the shared result was now in front of each one.
If you run several agents on one job, you can do a version of the same thing and check it. Give each the shared goal, not only its own piece. Then judge the run on the combined result, for example by running the full test suite on the merged code and reading the merged output yourself, rather than trusting each agent’s report that its part is done.
Two shifts, one pile
Imagine you manage a warehouse with two shifts, each scored on orders shipped per hour. The night shift’s number keeps climbing. The day lead keeps complaining that every morning starts with a cart of half-processed returns.
You add a second line to both leads’ scorecards: returns cleared, counted across both shifts. The next week the night shift’s orders per hour dip.
The morning cart is gone.