The Room Was Full Of Private Conversations

At the last on-site, we were all in the same room.

So were our agents.

You could not see them, of course, but you could feel the extra population. Each person had several threads running: one agent researching, another editing, another moving through a repository, another waiting for an answer. We had gathered in person and brought along a small invisible workforce that did not need chairs, coffee, or introductions.

The room felt more crowded than the headcount suggested. It also felt lonelier.

At earlier on-sites, work had a social soundtrack. Someone would turn a laptop around. A half-formed question would escape into the room. Two people would rubber-duck a problem badly until one of them heard the missing assumption. There were interruptions, jokes, short arguments, and the occasional useful sentence that belonged to nobody in particular.

This time, there was more prompting.

When I took a break and looked around, I mostly heard keyboards. Click-clack from one side of the room. A short pause. Faster click-clack from somewhere else. Everyone appeared intensely occupied inside a conversation the rest of us could not hear.

I felt guilty for being still.

That made no sense. I had multiple threads moving. Work was happening because I had already set it in motion. The useful thing for me to do in that moment may have been to wait, think, look up, or ask whether any of those threads deserved to exist. But the room offered an older measure of contribution: other people were typing, and I was not.

The feeling may be newer than the office, but the machinery behind it is not. In a study of supermarket workers, economists Alexandre Mas and Enrico Moretti found that people increased their effort around highly productive coworkers, especially when those coworkers could see them. A checkout shift is a long way from agentic knowledge work, but the social reflex is familiar. We read the effort around us and adjust our own. Visible activity becomes a quiet form of pressure.

Knowledge work has always made that signal unreliable. Thinking can look like staring out a window. Finding the dangerous assumption can look like taking a walk. Now agentic work makes the signal stranger still. A person can be orchestrating several active processes while sitting motionless, then type furiously because they are correcting work that should never have been delegated. The keyboard reports motion. It does not report value.

There is another old human mechanism in the room: proximity is not the same as community. Research on coworking spaces has found that people do not get a sense of belonging merely by sharing furniture and Wi-Fi. Community is something workers co-construct through ordinary, repeated interactions. The small talk, the request for help, the shared ritual, and the unplanned exchange are not decorations around the work. They are part of how a room becomes social.

Agents are very good at intercepting those little exchanges.

The question that once crossed the table now goes into a prompt. The rough idea that might have been tested against a coworker's expression goes into a private thread. Even rubber-ducking changes when the duck can answer, revise the code, search the documentation, and reassure you that your approach is reasonable. The convenience is real. So is the disappearance.

I was curious about the feeling, so I dug a little deeper: what changes between people when more of our working relationship is routed through a machine?

Some early AI research can already see parts of this. Across four studies involving 794 employees in Taiwan, Indonesia, the United States, and Malaysia, researchers found that more interaction with AI at work was associated with both loneliness and a greater need for social affiliation. That need sometimes led people to help coworkers more, even as loneliness carried costs after work. The paper's title, “No person is an island,” is almost too perfect for a room where every person has acquired a private archipelago of assistants.

A newer analysis of Microsoft 365 activity offers a different view of the same shift. Among heavy users in the study, generative AI adoption was associated with increases in both productivity-oriented and communication actions, but the balance moved toward more individual, document-focused activity. The researchers are careful about what digital traces can prove, but they warn that efficiency gains could still weaken interpersonal communication and the circulation of diverse information. The work can become more active and less shared at the same time.

That is not the only possible outcome. A randomized field experiment with 316 employees at a European technology services firm found that access to a company-grounded AI assistant increased connections in collaboration and knowledge-sharing networks over three months. In that setting, the “digital colleague” did not pull people apart. It changed who talked to whom and increased connectivity.

The research is young, and none of it can tell me what everyone else in our room felt. It can only make the scene less mysterious. AI does not carry loneliness into the office like a feature flag. It changes the cost of asking, explaining, drafting, waiting, and reaching across the table. Those small changes meet an older animal that watches other people work, mistakes movement for contribution, and needs interaction to turn physical presence into belonging.

An on-site used to create conversation partly by removing the effort required to begin one. The person was right there. Agentic work lowers the effort of a different conversation even further. The assistant is already inside the task. It has the file open. It does not look busy, require backstory, or make you wonder whether the question is worth interrupting someone for.

That means being together may need more intention than it used to.

Not another scheduled block called “human connection.” The calendar has done enough. Maybe it is as small as showing the messy middle, reviewing an agent's wrong turn together, saying the half-formed thought before converting it into a prompt, or asking the room a question that a machine could answer faster. The value would not only be a better answer. It would be making part of the work audible again.

I keep returning to the guilt I felt when I stopped typing. Several machines were working on my behalf, yet the room made stillness feel like absence. Maybe other people felt it too. Maybe they were typing partly because I was typing. We could have been amplifying the sound of productivity for one another while each privately wondered whether we were doing enough.

The agents did not invent that feeling. They gave it a strange new room to live in.

We were together. We were getting things done. We had more help and were getting more done than ever before. But for a few minutes, listening to all those keyboards, I could not tell whether we had gathered to work with one another or simply agreed to be alone in the same place.