Where Good Engineers Come From
An accidental experiment on how engineering skill moves between people, and the bill attached.
When Covid closed the offices in March 2020, the engineers of one large company went home like everyone else. Years later, economists studied that company’s code reviews, before the closure and after. The numbers show something most of us only suspected.
Engineers who sat physically close to their teammates used to get more comments on their code, and better ones. When the offices closed, that advantage fell by 18%. The feedback did not move into chat or calls; the researchers checked, and losing the office reduced the online conversations too. The desk conversations were never in the data at all, so the real loss is bigger than the number. The comments that teach disappeared faster than the ones that just fix style. Junior engineers felt the loss most.
The pattern goes beyond software. In German wage data, people who work near more knowledgeable colleagues see their pay grow faster for a decade afterward, and the economists estimate that 4% to 9% of total pay is really payment in learning.
There is a sadder version of the same evidence. Researchers followed scientists after the death of a senior colleague they had worked closely with. Their output fell by 5% to 8% and never fully returned. Part of what the senior knew was never written down. It lived in the conversations, and it left with them.
So far, this is the story every senior engineer likes to believe: the craft spreads like an infection. Juniors catch their seniors’ standards just by being near them, and workshops and KPIs are theater around that.
But the story has three problems.
First, formal training works better than the folklore says: in the largest reviews of training research, courses produce a solid, measurable change in how people behave on the job. The famous 70-20-10 rule, the one claiming only 10% of learning comes from courses, traces back to a 1988 survey that asked 191 successful executives to remember what shaped them. It is folklore with a citation.
Second, whether training sticks depends on what happens after the course. The strongest predictor of transfer is support from supervisors back on the job. A junior comes back and tries the new method. If their senior makes room for it, it becomes practice. If the senior shrugs, it dies in a week. Seniors decide whether the workshop was worth anything, usually without noticing that they are deciding.
Third, the way skill moves is less romantic than it looks. The cleanest peer study outside software watched supermarket cashiers, who worked faster when a productive colleague could physically see them: line of sight and social pressure. And in the office study, the learning traveled through code review comments: a tool, a queue, a process.
That covers the workshops. The KPIs deserve their reputation too, with a footnote. Once a metric becomes a target tied to rewards, effort moves to the measured work and quality moves from the unmeasured work; across more than a hundred experiments, rewards tied to performance reliably reduced people’s own drive to do the task well. The footnote: metrics read as diagnostics are a different thing from metrics used as targets. Even DORA warns against using its four numbers to grade teams.
So the belief is half right. Sitting near skilled seniors is a real, measurable channel for skill, probably stronger than a workshop and certainly stronger than a dashboard. But it is one channel inside a larger system. Code review, standards and hiring do heavy lifting too.
That channel is expensive. In the office study, experienced engineers wrote measurably less code while sitting near teammates. Mentoring showed up as a tax on their output. The tradeoff fell harder on women, who gave more mentorship and lost more of it when proximity disappeared.
The three problems connect here. The channel that carries the culture runs on senior attention, attention costs output, and output is what dashboards measure. So a company that manages seniors by individual numbers is quietly taxing its own teaching. The dashboard cannot see the curriculum. It only sees the cost.
The dashboard misses the risk side too. Knowledge that travels through proximity concentrates in a few heads, and when the person everyone learned from leaves, the code they held goes quiet and stays quiet. An apprenticeship culture copies everything, including the shortcuts and the blind spots, and it quietly underserves anyone who does not sit near, or socially resemble,the seniors.
Software already borrowed one habit from a field with the same problem. The blameless postmortem on your calendar came from aviation and from medicine’s case review meeting. In its home field, that meeting sits inside a funded system: Medicare alone sends US teaching hospitals more than $15 billion a year, so a senior doctor’s teaching hours come out of a budget instead of out of their own output. Medicine has its own problems, and its hierarchy is one of them. But it treats teaching as normal work. Software borrowed the meeting and skipped the budget.
Until your company builds that budget, the system runs on you. If you are the senior in this story, the real training program is the sum of your review comments: what you praise, what you let pass, who gets your attention. When your commit count falls while your juniors improve, that is the system working. Say it out loud before a dashboard says otherwise. If you run a team, put the teaching hours in the plan, with a number next to them. When someone comes back from a course, ask what they learned and make room to use it. And in a remote team, the review queue is the new office map. Decide who gets your comments, or the tool will decide for you. The people learning from you are mostly the ones who can see you, and that group is smaller, and less fair, than most of us assume.
Sources:
Emanuel, Harrington and Pallais, “The Power of Proximity to Coworkers,” Quarterly Journal of Economics (2026).
Jarosch, Oberfield and Rossi-Hansberg, “Learning from Coworkers,” Econometrica (2021).
Azoulay, Graff Zivin and Wang, “Superstar Extinction,” QJE (2010).
Arthur, Bennett, Edens and Bell, Journal of Applied Psychology (2003), behavioral effect d = 0.62.
Blume, Ford, Baldwin and Huang, “Transfer of Training,” Journal of Management (2010).
Mas and Moretti, “Peers at Work,” American Economic Review (2009).
Deci, Koestner and Ryan, Psychological Bulletin (1999).
Austin, Measuring and Managing Performance in Organizations (1996).
McCall, Lombardo and Morrison, The Lessons of Experience (1988), origin of the 70-20-10 rule.
Congressional Research Service, Medicare Graduate Medical Education (2025).
Sydor and colleagues, British Journal of Anaesthesia (2013), trainees speaking up to seniors.
Allspaw, Blameless PostMortems and a Just Culture (2012).
