For much of its history, Google ran engineering on a simple theory: leave people alone. If engineers got stuck, they could ask their bosses, whose deep technical expertise had put them in charge in the first place. In 2002 the founders had even tried removing engineering managers altogether. By Harvard Business Review’s account, the experiment lasted only a few months.
So when Google’s people analytics team began in 2009 to study what its best managers did, the company’s working belief was clear. Laszlo Bock, who ran people operations, put it plainly: “In the Google context, we’d always believed that to be a manager, particularly on the engineering side, you need to be as deep or deeper a technical expert than the people who work for you.”
The team, code-named Project Oxygen, coded more than 10,000 observations about managers and came up with eight behaviors. Then Bock’s group ranked them.
Technical expertise came last.
The pan you keep reaching for
It was important, Bock told The New York Times, but it paled next to “making that connection and being accessible.” Be a good coach. Don’t micromanage. Take an interest in people’s careers.
The skills that got you promoted are not the job you were promoted into.
For a book published in 1992, Linda Hill of Harvard Business School followed 19 new managers through their first year, and she kept studying star performers promoted to manage. She found they arrived with misconceptions that, in her words, “rise almost to the level of myth.” They expected the job to bring authority and freedom. Instead, they found themselves hemmed in by subordinates, bosses and peers. “The fact is that you really are not in control of anything,” one told her.
One of the managers she describes had run a trading desk at an investment bank. To establish his credibility with seasoned traders, he told them which positions to close and which strategies to try. They pushed back and wanted his reasoning for every call. Things changed when he asked one of them a simple question about pricing in a market he didn’t know. “Once I stopped talking all the time and began to listen,” he said, the desk “started to educate me about the job.”
The best line cook made head chef keeps reaching for the pan. The kitchen needs her tasting every plate, and sending some back.
The query you fixed yourself
Imagine you were the strongest analyst on your team, and now you lead it. A junior analyst’s dashboard query takes four minutes to run. You could explain indexes to her on Thursday. Instead you rewrite it yourself that night, and by morning it runs in six seconds.
She thanks you. The next week she sends you the next slow query, with a note: could you take a look?
Where expertise wins back
Project Oxygen is Google’s own account of Google, and HBR noted its limit: nobody could yet show how much the higher manager scores moved sales, productivity or profit.
Other evidence pulls the other way. Economists Benjamin Artz, Amanda Goodall and Andrew Oswald looked at US and British survey data on workers and their supervisors, using measures such as whether the boss had worked up inside the company and could do the worker’s job in an emergency. A boss’s technical competence was the single strongest predictor of a worker’s job satisfaction. They call their study “a cautious attempt” and say its correlations can’t match a randomized trial.
One reading fits both. Google’s engineering managers were mostly deep experts already, so expertise couldn’t tell good managers from bad ones there. Neither study tested that reading. On that reading, expertise is the ground a manager stands on, and the day goes to something else.
Managers of agents
In August 2025, Anthropic surveyed 132 of its own engineers and researchers and interviewed 53 of them about working with its AI model, Claude. Many described their role shifting from writing code to managing AIs. Some saw themselves as “manager[s] of AI agents.” One estimated their work had shifted “70%+ to being a code reviewer/reviser rather than a net-new code writer.” Several compared it to the move from individual contributor to manager. Yet more than half said they could “fully delegate” only 0 to 20 percent of their work.
Here technical skill matters again. Some worried about “skills atrophying” as they delegated more, and the report names a “paradox of supervision”: using Claude well requires supervising it, and supervising it requires the very coding skills that heavy use may erode. For a manager of agents, expertise may be the job.
The company says the findings likely don’t generalize: its staff are early adopters, the interviewees were the first 53 to respond, and the models were those of August 2025.
You can test your own footing. Take the last agent-written change you approved. Close the agent’s summary and explain the change to a colleague from the diff alone. Wherever you can’t, go back to the evidence. It may be a part you didn’t check, or one you’ve forgotten or never fully understood.