“One of the most valuable things I learned was to give the appearance of being courteous while withholding just enough information from colleagues to ensure they didn’t get ahead of me on the rankings.”
The engineer who said that worked at Microsoft. He was describing what the company’s review system taught him. Kurt Eichenwald collected it, along with accounts from other current and former employees, for a 2012 article in Vanity Fair about Microsoft’s decade under Steve Ballmer.
The system was called stack ranking. Every unit had to sort its people into fixed shares: top performers, good, average, below average and poor. Every six months, managers from a group gathered in a conference room with the blinds drawn and placed Post-it notes bearing their staff’s names into a grid. “If you were on a team of 10 people,” a former developer said, “you walked in the first day knowing that, no matter how good everyone was, two people were going to get a great review, seven were going to get mediocre reviews, and one was going to get a terrible review.” Then he named the effect. “It leads to employees focusing on competing with each other rather than competing with other companies.”
Under that curve, each person’s surest move was a local one. Even hitting every one of your objectives didn’t guarantee a high ranking, since someone else might exceed theirs. So, Eichenwald wrote, people tried to do a good job and also worked hard to make sure their colleagues did not. Executives told him that many of the company’s stars avoided working alongside other top developers, for fear of what it would do to their ranking. The company’s own staff surveys, taken every six months, kept returning the same message: groups that were supposed to be working together weren’t.
Each of those choices made sense for the person making it. Added up, they made a company that couldn’t cooperate with itself. The case has the rest.
The sum of the best parts
A local optimum is the best result for one part of a system, by that part’s own measure. The trap is assuming that if every part hits its local best, the whole will too. Often it doesn’t, because the parts depend on each other, and a measure that rewards one part can reward it for making things harder for the next.
The trap isn’t only about people competing. At a large public hospital in New Zealand, the pharmacy that made chemotherapy drugs waited for a specialist to confirm each dose before making it. That protected the pharmacy’s own efficiency, since the drugs were expensive to waste. It also left patients waiting hours for treatment. Researchers using the theory of constraints listed “global measures, rather than local (or silo-based) optima” among the remedies, and the hospital found a way to meet both needs. The theory of constraints page tells that story.
Two of W. Edwards Deming’s 14 points for management speak to the same trap: “break down barriers between departments,” and abolish “the annual or merit rating.”
What Microsoft changed
On November 12, 2013, Microsoft’s head of human resources, Lisa Brummel, told staff that the curve was gone. Her memo also changed what counted as good performance. Reviews would now look at “not just the work you do on your own, but also how you leverage input and ideas from others, and what you contribute to others’ success.” A forced distribution, she told GeekWire, “wasn’t getting at the teamwork principles that we really want to get at.”
That is the usual way out of a local optimum: measure each part, at least partly, by what it does for the whole.
The hotel
Imagine a small hotel. The housekeeping team is measured on rooms cleaned per hour, so it works down each corridor in order, which is fastest. The front desk is measured on check-in time. On a busy afternoon, guests arriving early wait in the lobby for rooms that will be cleaned last, because they happen to be at the end of a corridor. Both teams hit their numbers, and the lobby fills up. If housekeeping cleaned in the order guests were due to arrive, its rooms-per-hour figure might dip a little, and the hotel might work a lot better.
Agents that each do their part
The same trap shows up when AI agents split up work. Anthropic built its Research feature with a lead agent that hands pieces of a question to subagents, which search in parallel. Early on, the lead agent gave short instructions like “research the semiconductor shortage.” The subagents did their jobs, and the whole still went wrong. Those instructions, Anthropic wrote, “often were vague enough that subagents misinterpreted the task or performed the exact same searches as other agents.” In one run, a subagent looked into the 2021 automotive chip crisis while two others duplicated work on 2025 supply chains.
The fix came from the level of the whole. “Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries.” That asks more of the lead agent up front. It is the price of a division of labor in which each part’s best work adds up.
The Microsoft engineer had learned how to play his local game. The company needed him to be playing a different one.