Land the Plane

Software engineering, AI-assisted development, and what it actually takes to lead engineering teams in the agentic era.

Episode 002 · 2026-05-30 · 24:45

Crew Rest

Why time with other humans is the maintenance schedule for the part of the work AI can’t run.

Published 2026-05-30. The audio version of this piece is Land the Plane episode 2; this post covers the same ground for people who’d rather read.

This week on the radar

Four things from the last little while.

Engagement hit a fifteen-year low. Gallup’s State of the Global Workplace 2026, released in April, reports global employee engagement falling to 20% in 2025 — the lowest since 2020 — and pegs the cost of disengagement at roughly $10 trillion in lost productivity. Treat the dollar figure as directional, not precise; it’s a model, not a measurement. The number that matters more: more than one in five employees report feeling lonely at work, climbing toward 26% for workers under 35 (FlexOS summary). We’re going to spend most of this piece on that word, lonely — it’s more load-bearing than it looks.

A 233-study review on loneliness at work. A 2026 review in the Journal of Management, “All the Lonely People” (McCarthy, Erdogan, Bauer, Kudret & Campion), synthesizes 233 empirical studies and draws the distinction the whole piece hangs on: isolation is objective (being physically alone); loneliness is subjective (the sense that your relationships are deficient). You can be deeply lonely in a crowded office — or in a Slack with four hundred people in it. Proximity doesn’t fix it; connection does.

Edmondson on AI as a two-edged sword. Amy Edmondson — who put psychological safety on the map — has spent the last few months applying it to AI: “only in environments where people feel safe do they speak openly, experiment, and view AI as an opportunity rather than a threat.” The sharp question she keeps asking: will people feel safe admitting they don’t understand a recommendation the AI just made? Hold that — it’s the bridge back to last week.

The counter-data, for honesty. Anthropic’s 2026 Agentic Coding Trends Report complicates the easy “AI isolates engineers” story: developers now use AI in ~60% of their work but fully delegate only 0–20% of tasks — and the harder, more design-dependent the work, the more they keep humans in the loop. So the tool isn’t forcing isolation. The villain here isn’t the agent. It’s the temptation, and the option.

The week where nothing went wrong

Here’s a scene. It’s Thursday. You’ve had a great week, by one definition of great. You and your agent have been heads-down since Monday. You shipped a service. You cleared a month-old backlog. The diffs are beautiful, the tests are green. Every morning you sat down, described what you wanted, and watched it get built. The agent answered instantly, never pushed back, never needed anything, never had a bad day. By every measure available on your screen, it was the best colleague you’ve ever had.

Then you stand up to get a glass of water and notice something strange: you haven’t had a real conversation with another human since Monday. Not a standup. Not a thread. A conversation — the kind where somebody disagrees and you have to sit with it; the kind where you say I’m stuck and someone you trust says yeah, that part is genuinely hard, let me look. By the org chart you were fully connected all week. Calendar full, messages flowing. And you were, in the precise clinical sense from that Journal of Management review, lonely.

Here’s what makes it dangerous: nothing went wrong. The week worked. By every dashboard your manager looks at, you crushed it. The cost didn’t show up anywhere a dashboard can see — it showed up as a slow, quiet draw-down on something that doesn’t have a meter on it.

This piece is about that meter you can’t see. It’s about why time away from the keyboard, with other humans, isn’t a reward you earn after the work — it’s an input to it. And it’s about a finding that genuinely unsettled me: the people most likely to be lonely in your org aren’t the disengaged stragglers. They’re the ones at the center.

The title is “Crew Rest.” Here’s why.

Rest is a maintenance schedule, not a reward

In aviation, crew rest is a hard, legally mandated rest period between duty. You can’t waive it by being a hero, can’t push through because the schedule’s tight. The rule exists because the industry learned, expensively, that a depleted pilot is a hazard to everyone on the plane — and that the pilot is the worst-positioned person to judge their own depletion. So the rest isn’t discretionary. It’s on the maintenance schedule, next to the engine inspections. It’s a safety system, because it is one.

Put engineering rest in that category. Not a perk, not a thing you earn after suffering enough — a maintenance system that keeps the most expensive component, human judgment, from failing in a way nobody catches until 2 AM.

And here’s where the science gets more specific than generic take-a-break advice. A body of recovery research — much of it descending from Sabine Sonnentag — asks: what actually makes rest restorative? The answer isn’t “time off the clock.” Plenty of people take time off and come back just as fried. The two ingredients that consistently drive recovery are psychological detachment — actually mentally leaving the work, not just physically — and positive emotion.

Sit with detachment. The trouble with how a lot of us rest now is we never do it. We close the laptop and keep running the bug in the background. We’re at dinner and cognitively still in the codebase. The agentic workflow makes this worse, because the agent is always available — there’s no natural stopping point, the build is never done-done, you can always kick off one more run before bed. The friction that used to force detachment — the office closing, the twenty-minute compile that sent you for coffee with somebody — much of it is gone. The agent removed it. So detachment is no longer something that happens to you. It’s something you have to choose.

This is where rest meets the rest of the piece. The APA’s framing of seven types of rest names social rest — restoration with supportive others — as a distinct category from physical and mental rest. You can’t sleep your way to it or get it from a quiet room. You get it from people who refill you instead of drawing on you.

But — and the research is clear, I don’t want to oversell it — social time only restores you if you’re actually detached while you’re in it. Lunch with a coworker where you grind through the incident the whole time isn’t social rest; it’s a meeting with worse posture. The recovery comes from the combination: with other people, and genuinely off the work.

The sabbatical research adds a timing finding I love: the people who got the most out of extended leave weren’t the ones who collapsed into it half-dead. They rested proactively, before depletion, and in emotional and relational modes — time with people they loved — not just sleep. The benefit persisted for months after they returned. Rest done right isn’t a debt you pay down. It’s a capability you build.

So the first move is to reclassify the thing. Time away, with other humans, isn’t the reward for the work. It’s part of the work. It’s crew rest. And in a world where the agent will happily let you fly twenty hours a day, you’re the only one who can ground the plane.

Relationships are the account you draw down on

Rest with people refills the person. Now: what it builds between people — the part that shows up in your engineering outcomes.

Start with the most-quoted study in the space. Google’s Project Aristotle studied ~180 teams across 250 attributes over two years to answer one question: what makes a team effective? They expected it to be the mix of talent. It wasn’t. The number one differentiator, above all of it, was psychological safety — the shared belief that the team is safe to take risks in, to be wrong in, to say I don’t know in. The line from the writeup I keep coming back to: “no one wants to put on a ‘work face’ when they get to the office.” The teams that won were the ones where people could show up as people.

Psychological safety is the outcome. For the input, reach for Patrick Lencioni’s Five Dysfunctions of a Team, which puts trust at the base of the pyramid — everything else (healthy conflict, commitment, accountability, results) sits on top of it. His definition is the part to read carefully: a willingness to be “completely vulnerable with one another — to let down our guard, admit our flaws, and ask for help.”

Admit our flaws. And ask for help. Hold those against the agentic workflow and something quietly alarming happens. Asking for help used to be a social act — when you were stuck, you turned to a human, exposed that you didn’t know the thing, and asked. That little moment of vulnerability, multiplied across a thousand small stuck-moments a year, is a huge amount of how trust actually gets built. You ask, they help, you’re slightly indebted and slightly closer, and next week it goes the other way. That’s the flywheel.

The agent removes the moment. Now when you’re stuck, you turn to the model. It answers instantly, never makes you feel dumb, never remembers you asked the same thing yesterday. Honestly a better experience than a busy senior who sighs before they help. So all of us are going to ask the agent instead — thousands of times a year — and every one of those is a deposit into the trust account that no longer gets made. That’s the muscle that atrophies: not because anyone decided relationships don’t matter, but because the single most common occasion for building one — I’m stuck, can you help — got quietly intercepted by a tool that does it faster.

This isn’t only about feelings. Stephen M.R. Covey’s The Speed of Trust makes the economic claim cleanly: “When trust goes down, speed will also go down and costs will go up. When trust goes up, speed will also go up and costs will go down.” Trust is a literal multiplier on how fast a group can move, because high-trust teams don’t have to verify everything, document defensively, route everything through process, cover themselves at every step.

And here’s the agentic irony I can’t stop turning over. Agents made typing nearly free — they collapsed the cost of producing code to almost nothing. But the speed of everything that happens between people — agreeing on what to build, trusting a teammate’s judgment, accepting hard feedback, deciding to kill a project — is still entirely governed by trust. We optimized the cheap part to zero and left the expensive part exactly where it was. Proportionally, the human part is now almost the entire cost. The relationship is the bottleneck now — not because it got slower, but because everything around it got instant.

So the time you invest in the people you work with isn’t extracurricular. It’s you funding the account the whole team draws down on every time something is genuinely hard — every honest code review, every this-is-not-good-enough, every I-think-we’re- building-the-wrong-thing. Those withdrawals are only possible if someone made the deposits. And the deposits get made off the keyboard: in the lunch, on the walk, in the unhurried hour where you’re two humans, not two roles.

The loneliness accelerant

Before the finding that got under my skin, don’t file loneliness under “soft” the way our industry files most things about humans — because there’s a hard number on it. In 2023 the U.S. Surgeon General put out an 81-page advisory, Our Epidemic of Loneliness and Isolation, with an anchor finding you can’t unhear: the mortality risk of chronic social disconnection is comparable to smoking up to 15 cigarettes a day — roughly 29% higher heart-disease risk and 32% higher stroke risk — and it names workplaces directly as a domain where “performance, productivity, and engagement are diminished.” We spent two decades on ergonomic chairs and standing desks to protect engineers from back pain, and put a fruit bowl in the kitchen for a relational risk on the order of a pack-a-day habit. The layer we treat as soft and optional is, by the numbers, one of the most dangerous things in the building.

Now the finding that inverts what you’d assume.

You’d assume loneliness at work concentrates at the bottom — the disengaged, the checked-out. It doesn’t. Gallup’s 2026 data says leaders report the highest engagement (26%) and the highest thriving (43%) — and the highest loneliness (31%) (Gallup). The people at the center of the network — most connected by org chart, in every meeting, copied on everything — are the loneliest in the building.

Read that next to the isolation-vs-loneliness distinction and it snaps into focus. Loneliness isn’t about how many people are near you; it’s about whether the connections are real. The people at the center have the most connections and the fewest real ones, because every interaction has an agenda attached — everyone wants something, or is performing, or is managing up. The org chart connects you to everyone and bonds you to no one.

Now drop an agent into that life. Here’s the trap, stated as plainly as I can:

An agent gives you the sensation of collaboration while delivering none of its value.

Think about what a real collaborator does. They answer your question — yes. But they also push back. They also need something from you sometimes. They remember you were short with them last week. They catch the thing you got wrong, because they have skin in the game and a reputation on the line. They occasionally just want to talk to you because they like you. The friction and the need and the memory and the reciprocity aren’t bugs in human collaboration — they’re the part that builds the trust, and the part that catches the error.

The agent has none of it. Instant, never pushes back socially, never needs anything, never remembers a slight. Pure frictionless responsiveness. And that feels wonderful — like the best colleague you ever had. But it’s the exact profile of an interaction that feels deeply social and is, functionally, completely solitary. It scratches the itch of collaboration without doing any of the things collaboration is for.

That’s what I mean by: AI is a loneliness accelerant disguised as a collaborator. Not that it isolates you by force — the Anthropic data says people keep humans in the loop on the hard stuff. It’s subtler and worse than force: it removes the felt need for other people. It satisfies the surface symptom of wanting company while the underlying thing — real connection, real trust, real relationship — quietly starves. You don’t notice you’re starving because you feel full. Hunger you’d fix. Fullness you ignore.

And here’s where it becomes an engineering problem, which is the through-line of this show. Last week the argument was that psychological safety is the critical path — catching bad agent output requires a team where people can say this is wrong without being punished. This week is the foundation underneath that: you cannot have psychological safety on a team with no relationships. Safety isn’t a policy you announce or a value on a wall. It grows in the soil of actual human trust, built over actual human time. You can’t install it. You have to grow it. And it grows off the keyboard.

Run the two episodes together and the dependency chain is stark. The agent generates plausible, confident, sometimes-wrong output at ten times the old rate. Catching it requires someone with judgment to say so out loud. Saying it out loud requires psychological safety. Safety requires trust. Trust requires relationship. Relationship requires unhurried human time — the exact thing the agentic workflow quietly deletes, because the agent removed every natural occasion for it. The chain runs straight from your 2 AM incident all the way back to whether you ever go to lunch with anybody anymore.

That’s the argument. The scarce input is no longer output. It’s the relational capital between people — the one input on the whole system you can’t buy more compute to fix.

In the agentic era, time with other humans isn’t the thing you do instead of the work. It’s the maintenance schedule for the one part of the work AI can’t run.

What to do this week

Three things, if any of this lands.

  1. Take one piece of rest with real detachment in it — and make it social. Not an evening doom-scrolling alone; that’s not rest, it’s a different screen. One meal, walk, or coffee with a human where the deal is no shop talk and no checking the build. Detached and with people — that’s the combination that refills you. Treat it like an engine inspection. Put it on the schedule. Don’t skip it because the sprint is tight. The sprint is always tight.

  2. Take one stuck-moment to a person instead of the agent. Not because the agent would answer worse — it might answer better. Do it because the point was never just the answer; it’s the deposit. You ask, they help, the account grows. Pick the colleague you’d like to trust more in six months and go be slightly, deliberately vulnerable with them today. You have to use the muscle on purpose now, because the workflow won’t make you anymore.

  3. If you’re at the center of the network, assume you’re at risk. Lots of meetings, lots of threads, copied on everything — that’s exactly the profile Gallup found loneliest, and your most senior people fit it too. Being central isn’t being connected. Make one real connection this week with no agenda — nothing you want from them — just time with a person you respect, because the relationship is the point. The most connected people in your org are the ones starving, and they’re the ones whose judgment your hardest decisions depend on. Feed that.

Go back to the cold open — the week where everything shipped, nothing went wrong, and you still ended up depleted and alone. The cost that didn’t show up on any dashboard. That’s the most dangerous kind there is: the kind the system can’t see and therefore can’t defend against. You’re the only instrument that reads that meter. So read it.

The agent will let you fly forever. It will never tell you to land. That part is on you.

Land the plane. And then, for the love of everything, go get lunch with somebody.

Sources

← all episodes