The Team You Never Meet

A co-located team is running a quantity of infrastructure that nobody installed and nobody pays for. Information moves through it constantly: an argument two desks away that tells you a decision is being made badly before it is made, a face that says the estimate was optimistic, a question asked sideways because asking it formally would have felt like escalating. None of that is on an organization chart. All of it is load-bearing, and it is invisible in exactly the way things are invisible when they are free.

When teams went distributed, the parts that had names got replaced first, because things with names are easy to procure. Meetings became video calls. Standups became standups. The document repository moved and the chat tool grew channels. Every one of those replacements was a line item somebody could approve, and every one of them was a replacement for the visible half.

That is the whole argument, and I want to state it before complicating it: the teams that struggle are not the ones with bad video calls. They are the ones that replaced the half with names and never noticed the other half was gone. What follows is what I think actually breaks, and then what helps, with an honest note on what each thing costs, because most of the advice in this area is free to give and expensive to take.

Onboarding is the clearest case and has the worst consequences.Learning by proximity is the largest thing that disappeared and the least replaced. Emanuel, Harrington and Pallais studied software engineers at one large firm across the 2020 closures and the later return mandates, in a 2023 paper, and found sitting near teammates raised the feedback engineers received on their code by about 18 percent, concentrated among the less-tenured. It is a natural experiment rather than a randomized one, at one firm, with feedback measured as comments in a review tool. It is still the closest thing to a measurement of the effect everybody asserts.

Weak ties go first, and they are the ones carrying the information.People keep the five relationships they need for their work and lose everybody else. Granovetter's 1973 argument was that novel information travels along weak ties precisely because close contacts already know what you know, and the evidence he leaned on was job-search interviews with professional men in one Massachusetts suburb, a much narrower base than half a century of citation suggests. The mechanism still describes what teams report: the close pairs survive, the far edges do not, and the cross-team knowledge that used to make things fast is the part that quietly goes.

And there is one large measurement of exactly that, with a large caveat attached.Yang and colleagues published a 2021 study of tens of thousands of employees at one company across the shift to firm-wide remote work, and found the collaboration network became more static and more siloed, with fewer connections bridging separate groups. The caveat is doing real work: it is one firm's communication metadata over the first six months of a pandemic, which is not a controlled comparison of remote against office, and it measures who talked to whom rather than whether anything got worse.

Interrupting costs more, so people wait.Asking a question used to cost a glance to see whether somebody looked busy. Now it costs a message that arrives without context, will be read at an unknown time, and might be interpreted as escalation. So people wait. They wait through the small confusion that would have been resolved in eight seconds, and the confusion is still there a week later, having grown into a design decision. The most expensive property of a distributed team is that a two-minute question and a formal request look identical when they arrive.

A distributed team has to write, and writing is better than talking for anything that has to outlive the conversation. A decision made in a room exists in the memories of the people who were there, degrades within weeks, and is unavailable to anybody who joins later. A decision written down with its reasoning is available to a person who arrives in two years and has no way to ask. That asymmetry is real and it is the one thing the arrangement hands you for free.

The version that works is duller than it sounds. A short written record per decision: what we chose, what we did not choose, what we believed at the time that made the difference, and what would have to change for this to be wrong. That last clause is the one that turns a document into something a future reader can act on. It takes twenty minutes and it is the twenty minutes that is always available for the meeting and never for the write-up.

And the honest cost. Writing well is slower than talking, it is a skill unevenly distributed across any team, and it disadvantages people who think out loud and people writing in a second language. A team that decides everything in documents has quietly changed who is influential. I do not have a clean answer to that. The partial one is to separate the thinking from the record, letting the argument happen however it happens and requiring only that somebody writes down where it landed and why.

The failure mode is not too few meetings.It is running the same meeting, slower. A question is posted, three people respond over two days, nobody has authority to close it, and the thread continues until somebody schedules a call, which was the outcome anybody could have predicted on the first day. That is not asynchronous working. That is a meeting with a worse interface and a longer duration, and it is what most teams get when they are told to reduce meetings.

An async decision needs a deadline and a decider, or it is just delay.The shape that works: a written proposal, a named person who will decide, a date the input window closes, and a default that takes effect if nobody objects. All four are necessary. Without the date, the discussion expands to fill whatever time exists. Without the default, silence blocks rather than consents, and the people who are busiest end up with the most power to stall things by saying nothing.

Pairing is the highest-fidelity replacement for learning by proximity, and it costs two people's hours to produce one person's output for the duration. There is no version where that is free, and the argument that it pays for itself in fewer defects is one I would not lean on hard. The reason to do it is narrower: it transmits the reasoning behind decisions, which is exactly what does not survive in a document, and it is the fastest way to get a new engineer to useful.

In-person time is the other purchase, and most organizations buy it badly. Flying people across a continent to sit in a room attending video calls and answering email pays the entire cost of gathering and collects none of the benefit. What genuinely goes better with everybody in one place is what needs trust and bandwidth: planning where the disagreements are real, the argument nobody wants to have in writing, the first week of somebody's job, and meeting the person whose name you have only seen on review comments.

None of this is cheap, and I would rather say that plainly than pretend there is a set of practices that recovers what proximity provided at no cost. The honest position is that co-located teams were getting something for free that distributed teams have to buy, and the choice is to buy it deliberately, buy it badly, or not buy it and absorb the consequences somewhere less visible.

You are now the deliberate substitute for ambient information.In a room, a manager absorbs the state of a team continuously and mostly without effort. Distributed, none of that arrives unless somebody builds a mechanism to carry it, and there is no natural moment at which its absence announces itself. If you have not built one, you do not have the information; you have the pleasant impression of having it, assembled from status updates written by people who want to sound in control.

Introduce people to each other on purpose, because nothing else will.Weak ties in an office formed by accident: shared lunch, adjacent desks, the same elevator at the same time. Distributed, the accidents stop and the network hardens into whoever was already connected. Deliberately connecting two people who ought to know each other feels artificial and is one of the higher-leverage things a manager does now, because otherwise a team of a dozen becomes six pairs that route everything through you.

Remote work works. The best evidence available is a randomized trial by Bloom, Han and Liang published in Nature in June 2024, which assigned more than sixteen hundred employees at a Chinese technology company to a hybrid schedule or to full office attendance and found quit rates fell by about a third with no damage to performance ratings or promotions over the following two years. That is a real experiment with a real control group, which is more than almost anything else in this literature has.

It also does not contradict a word of what I have argued, and I want to be careful not to claim it does. It tested two days a week at home at one company, which is a hybrid arrangement rather than a fully distributed team, and it measured attrition and performance ratings rather than whether a junior engineer became senior or whether the team knew what the team next to it was doing. What it establishes is that the productivity panic was wrong, which is a different question from mine.

A good deal of what gets mourned about offices was cheap sociability rather than productivity, and I have sat through enough performative presence to know the difference. The people who gain most from distributed work are carers, disabled engineers, people with long commutes and people who do not live near a technology hub, and they are consistently the ones not in the room when the policy is decided. Any argument in this area that ignores that is deciding for other people using its own convenience as the measure.

So I am not arguing for the office and I would not run a team any other way now. The claim that survives is smaller and more annoying: the accidental infrastructure was real, it was doing work nobody costed, and the version you build to replace it will be worse and will have to be maintained by hand. That is not an argument against distributed work. It is the bill for it, and most organizations have not opened it.

I have got this wrong in both directions: once mistaking a quiet channel for a calm team and learning otherwise at an exit interview, and once flying people to a room where they sat on video calls, which I told myself was about culture. The pipeline half of this, where a generation of juniors is harder to grow, is its own problem and I wrote about it recently. What I have never managed is a distributed team whose ambient information matched a room's, and I have stopped expecting to.

What I keep is a small and unglamorous rule: assume nothing arrives by accident. Every piece of information that used to reach me because I was standing near it now reaches me because somebody decided to send it, or it does not reach me at all. A team you never meet is not a team with worse tools. It is a team where everything that used to be overheard now has to be said out loud, on purpose, by somebody who thought to say it.