Late in 2001, with the dot-com bubble burst and the September 11 attacks behind it, Netflix, then about 120 employees, put its plan to go public on hold and laid off a third of the staff. “It was brutal,” wrote Patty McCord, the company’s chief talent officer.
That Christmas, DVD players turned out to be the hot gift. By early 2002 the DVD-by-mail business was growing fast, and Netflix had far more work and 30% fewer people.
One of its best engineers, whom McCord calls John, had managed three engineers before the layoffs. Now he was a one-man department working very long hours. McCord told him she hoped to hire some help for him soon.
He said there was no rush. “I’m happier now.”
Fixing other people’s mistakes
The engineers who had gone, he explained, weren’t spectacular. They were merely adequate, and he had spent too much time riding herd on them and fixing their mistakes. “I’ve learned that I’d rather work by myself than with subpar performers,” he told her.
Three adequate colleagues had not added three engineers’ worth of work. In his account, they had taken a share of his.
Adding people who need carrying can cost a team more than they add.
His words, McCord wrote in Harvard Business Review in 2014, came back to her whenever she described the most basic element of Netflix’s talent philosophy: the best thing you can do for employees is hire only “A” players to work alongside them. “Excellent colleagues trump everything else.”
In a shared kitchen, the best cook can spend the evening scrubbing other people’s pans. Take away the extra cooks and the meal may arrive sooner.
Reed Hastings and Erin Meyer built their 2020 book, No Rules Rules, around the idea under the name talent density. In Meyer’s telling, increasing it was the first of three steps Netflix took, before increasing candor and removing controls. Meyer told a 2021 business forum, in its organizers’ recap, that after the layoffs the remaining 80 employees were accomplishing almost double what the 120 had done before. Netflix’s culture memo now says a high performer in any role is “many times more effective than the average employee,” and that the company pays top of market to recruit and keep them.
Stars don’t always travel
The story has a problem built in. The layoffs landed just as demand took off, and the account comes from the people who ran the company. One firm’s good years can’t show how much was the smaller team and how much was the Christmas DVD boom.
Research on star performers pulls in a different direction. Boris Groysberg and colleagues at Harvard Business School followed more than 1,000 star stock analysts. When a star moved to another firm, they reported in HBR in 2004, her performance plunged, and so did the effectiveness of the group she joined. A later study found the decline lasted at least five years. It was worst for stars who moved alone or to weaker firms. Those who moved with their team showed no significant decline.
The authors argued that a star’s skills include company-specific ones, and that many of the resources behind a star’s results stay behind when she leaves. Read against Netflix, that is a warning about where talent lives. McCord’s engineer was judged at Netflix, inside Netflix’s systems; nobody measured him anywhere else. Density counted one person at a time can miss what a team supplies.
The colleague who needs carrying
Say you lead a six-person data team. One analyst is pleasant and willing, but most of his reports come back to you with errors, and two senior people now check his work before anything ships.
The Netflix culture deck’s answer, for someone you wouldn’t fight to keep, is a generous severance package. The density view at least makes the hidden cost visible: count the hours your two best people spend on his output. That count might also point elsewhere. If the errors come from a data pipeline nobody explained to him, removing him leaves the pipeline. (The keeper test asks a sharper version of the question, and carries its own costs in fear.)
Bigger model, smaller crowd
Teams of AI agents raise the same arithmetic, and Anthropic has published numbers on it. In its multi-agent research system, a lead agent hands parts of a question to subagents that search in parallel. Early versions, its engineers wrote in 2025, made errors like spawning 50 subagents for a simple query and distracting each other with excessive updates.
On one browsing benchmark, three factors explained 95% of the variance in performance: how many tokens were spent, the number of tool calls, and which model was used. More tokens helped most. But “upgrading to Claude Sonnet 4 is a larger performance gain than doubling the token budget on Claude Sonnet 3.7.” A better model beat doubling the older model’s token budget.
The parallel stops at morale. Nobody in that system was happier working alone. It was one benchmark inside one company, too, so it says nothing about people. What it adds is a measurement the Netflix story lacks. Swapping in a stronger model and giving a weaker one more tokens were run against the same test, and the stronger model came out ahead.