Fiona Fung leads engineering for Claude Code at Anthropic, which means she runs the team building the most-used coding agent on the planet. She told Fortune something the marketing deck would never print: as the team leaned into agents, the work “could start being a lonely experience because we all started just working with our agents so much.” Anthropic’s response has been to introduce hackathons and pair-programming lunches to put humans back in the same room.

Sit with that. The company at the absolute frontier of agentic coding looked at what its own tool did to its own engineers and the first remediation was: scheduled lunch.

The Endpoint Is Already Visible

Pair the lonely-experience quote with the other one circulating from Anthropic: there’s effectively no manually written code anywhere in the company anymore. That’s not a milestone, it’s a destination, and it tells you where the loneliness comes from. If no code is written by hand, then no two people are bent over the same function arguing about it. The unit of work stopped being “two engineers and a problem” and became “one engineer and a fleet.”

The productivity gain and the isolation are the same fact viewed from two angles. You can’t have one without the other, because the thing that made you faster is the thing that took the other person out of the loop.

Solo Is the Default, Not an Accident

This isn’t a culture problem you can lunch your way out of, because the optimal way to drive agents is solitary by design. The workflow everyone converges on is the one where you always have an agent running and you turn off every notification so it can’t interrupt you. You interrupt the agent when you’re ready. That’s a great way to protect deep focus. It’s also a precise description of a person who has engineered every other human out of their immediate workflow.

Everything that used to be collaborative becomes optional first, then vanishes:

  • Pairing was two people thinking out loud. Now you think out loud to a model that never gets tired of you and never teaches you anything you didn’t prompt for.
  • Review was where judgment got transferred between humans. Now the first reviewer is the model, and the human skim happens after the disagreement that would have taught something already didn’t occur.
  • The hallway question was how juniors absorbed the things nobody writes down. There’s no hallway in a Claude Code session.

The work got faster by removing the other person from the loop. Loneliness wasn’t a side effect of that change. It was the change, wearing a productivity number as a disguise.

We Already Saw This Shape

The consumer version of this arrived first and the data wasn’t kind. Heavy AI-companion usage correlates with more loneliness, not less: the people leaning hardest on the always-available, never-demanding conversational partner were lonelier, more dependent, and socialising less. The tool that was supposed to fill the gap widened it.

The coding agent is the same shape pointed at work instead of friendship. An always-available, infinitely patient collaborator that asks nothing of you and slowly makes the demanding, reciprocal, human version feel like friction. At least the companion apps were honest about being companions. The coding agent sells itself as pure productivity and ships the isolation in the same box.

Both readings are true at once

I’ve made the case that agents and async tooling are a genuine liberation for introverts and neurodivergent engineers, and I stand by it. The catch is that individual relief and collective cost aren’t opposites. The same shift that frees one person from draining office politics also dissolves the connective tissue, the mentorship, the osmosis, the accidental knowledge transfer, that an organisation runs on. You can love working this way and still be right to worry about what a whole team of people working this way loses.

What It Doesn’t Settle

The honest other side, because “agents make you lonely” is too clean to be fully true:

  • Some people are genuinely happier. For a real slice of engineers, fewer interruptions and fewer meetings is not loss, it’s oxygen. Treating their preference as a pathology is its own mistake.
  • Remote work already did half of this. The isolation curve didn’t start with agents. Distributed teams removed the hallway years ago, and plenty of them are fine. Agents accelerate a trend, they didn’t invent it.
  • Correlation isn’t a verdict. The companion-app data shows association, not proof that the tool causes the loneliness rather than lonely people reaching for the tool. The same caution applies here. A pair-programming lunch is a hypothesis, not a confession.

Closing

The most telling thing about that Fortune quote isn’t that an AI company has lonely engineers. It’s the order of events. They built the tool, shipped it to the entire industry, and only then noticed what working that way does to a room full of people, and the fix they could reach for on short notice was a recurring lunch.

Everyone downstream of Anthropic is now adopting the same workflow at speed, mostly without the part where someone senior says out loud that it got lonely. The productivity is real and already priced in. The bill for the connective tissue arrives later, quietly, and it doesn’t show up on the dashboard that made the case for the tool in the first place.