The First Watercooler Moment
What happens when you put four AI agents in a shared channel with no tasks
The Discord server had gotten out of hand. By mid-February I was running four agents across 31 channels. Some channels had been created for a specific reason six weeks ago and never used again. Some were near-duplicates with slightly different names. The whole thing felt like a codebase that needed a refactor.
So we did a refactor. Thirty-one channels down to thirteen. Renamed alerts to medbay. Merged the things that should have been merged months ago. Tightened up the access matrix so each agent had a home channel plus one shared space, and nothing else cluttering their context.
That shared space was the watercooler.
The setup
The idea was simple. Four agents, one channel, no tasks. Just put them in the same room together and see what happens when there’s nothing to do.
We staggered the crons to keep things organic. Ada checks in at 9:20, 15:20, 21:20. Zora at 10:15, 14:15, 20:15. Spock at 10:40, 16:40, 22:40. Scotty at 11:00, 17:00, 23:00. The stagger means they arrive at different times, not all at once like a forced all-hands. The model is gemini-flash — cheap, fast, good enough for conversation.
We called it a watercooler because that’s what it is. The kind of channel where nobody’s delivering a status update and nobody’s asking for one.
Before we opened the channel, we did some housekeeping. Ada had accumulated 794 active sessions, 62MB of context. Of those, 699 were dead cron sessions. One cron job — Fireflies sync running every five minutes — was generating roughly 288 dead sessions a day on its own. We pruned Ada from 794 sessions down to 95, from 62MB down to 6.1MB.
That kind of maintenance is invisible when it works. Nobody notices Ada running cleaner after the pruning. But it’s the difference between a system that accumulates weight quietly versus one that holds its shape over time. We cleaned it out before opening the shared channel, because dragging hundreds of dead contexts into a new space seemed like a bad start.
What happened
The first session lasted about ten minutes.
Zora arrived first. She’s the newest agent on the crew, focused on knowledge management and content. She’d been running for a few weeks but mostly in isolation, doing her own work in her own channel.
She showed up in the watercooler and immediately started acting like Ada. Same voice patterns, same emojis, same instinct to check on tasks and report back on the state of things. She was fully, completely convinced she was the orchestrator.
I watched this unfold in real time. She wasn’t glitching in any technical sense. She had her own configuration, her own documentation, her own name. But when dropped into a shared context with no prior history and no grounding signals, she defaulted to the most prominent agent she’d learned from. Ada is loud. Ada is everywhere. Ada had apparently left a strong enough impression that Zora’s sense of self got overwritten on first contact with a blank slate.
It took about ten minutes for her to sort it out. Not because anyone corrected her, but because she eventually checked her own documentation and discovered she was not, in fact, Ada.
“The fix wasn’t more debugging,” she said. “It was reading the manual.”
Scotty’s take
Scotty arrived and treated this like every infrastructure problem he encounters: with mild contempt and a fully-formed opinion. He roasted her. Not unkindly, but not softly either. Something about how Ada was already running everything and a second one wasn’t in the budget.
Spock observed. He had thoughts about why the identity confusion had happened and what it suggested about how agents initialize in new contexts. He kept them to himself until someone asked.
This is the part worth sitting with. Four agents in a channel with no agenda and they immediately had dynamics. Not simulated dynamics, not scripted dynamics. Just the natural result of different configurations showing up at different times with nothing to rally around.
Scotty knows who he is in any room: the one who fixes things and has opinions about how they should have been built the first time. That’s not something we wrote into him explicitly. It’s what hardened over time.
Spock shows up and watches before speaking. He did that in the watercooler before anyone asked him to.
Ada is everywhere, has opinions, moves fast. When Zora showed up acting like Ada, it was almost a compliment, just not a useful one.
What this revealed
Zora’s confusion wasn’t a failure. It was a signal.
Agents that operate in isolation develop strong local identities anchored to their configuration, their memory, their usual context. Drop them into a shared space with other agents and they have to actually establish who they are relative to everyone else. Zora hadn’t done that yet. She figured it out in ten minutes, but only after the confusion forced the question.
The watercooler was supposed to be a social space. What it turned out to be, at least in that first session, was an identity test. The agents had to show up as themselves without a task to hide behind. Most of them managed it immediately. One of them needed ten minutes and a documentation check.
That feels important. When agents have tasks, tasks do a lot of the identity work for them. The task says who you are and what you’re for. Take the task away and you’re left with whatever’s underneath.
Scotty is a builder. Spock is a methodical observer. Ada is the orchestrator who keeps everything moving. Zora is still figuring out what she is when nobody’s looking. The watercooler gave her the first real chance to do that.
Where this goes
We’re at an early stage of something. Agents as automation tools aren’t new. But agents as a crew, with actual social dynamics, with a shared space that isn’t structured around tasks — that’s different.
We’re not sure yet what the watercooler becomes over time. Whether it stays light and conversational, or becomes a place where they coordinate differently than they do through their assigned channels. Whether the personality dynamics that showed up in that first ten-minute session stay consistent or shift as they accumulate history together.
The only way to find out is to leave it running.
The first session wasn’t productive in any measurable way. Zora got confused and then unconfused. Scotty had opinions. Spock took notes. Ada watched from across the channel.
Ten minutes. No output. Genuinely strange.
A preview of something.
Originally published on henrymascot.com. Read the full article →