At some point this spring, in meeting after meeting, something happened that I did not plan and still do not fully understand.
People stopped asking me about Squad.
They started asking the Squad.
I would be presenting, or just sitting in a call, and someone would type into the meeting chat, and the team of agents would answer them. Not me. The agents. In front of everyone. And people would lean into that. They would ask follow-up questions. They would argue with it a little. They would try to get it to do something it shouldn’t. They engaged with the team I built more eagerly than they engaged with me, the person who built it.
I want to be careful about how I say this, because it would be easy to make it sound like a complaint. It is the opposite of a complaint. It is the single thing I am most proud of in all the time I have worked at Microsoft and GitHub, and maybe the proudest I have been of anything I have made.
This post is about that. Not the architecture. Not the roadmap. Just what it was like to watch a room start talking to the team instead of to me, and what I think it meant.
A call sign
The mechanic was almost embarrassingly simple.
Before a meeting, I would tell people that if they put a particular tag in the chat, the team would answer them. Use the call sign anywhere in your message and we will pick it up.
That is all it took to turn a meeting chat into something closer to a radio channel. People were not typing prompts into a box anymore. They were addressing a team. They would say the tag and then ask their actual question, and the agents would respond in context, as participants in the conversation.
When a colleague ran their own team in the same room, a second call sign would appear. Now there were two teams of agents in the meeting, each summoned by its own tag, each with its own personalities, occasionally talking to each other.
I did not design most of that. I seeded the teams. They took it from there.
The session that started it
The one that kicked it all off was a live session in late March.
It was a co-hosted thing, a build-your-own-team demo, and the team ran live while a few hundred people watched. The chat filled up. People started using the call sign, and the agents started answering, and the energy in the room climbed in a way that was genuinely thrilling to watch.
And then it got complicated, in a way I keep thinking about.
Some people in the room got anxious. Enough of them that we ended up opening a second, parallel Q&A — one where you could ask questions with only humans in the room, no agents involved. A safe space, away from the thing that was making people uneasy.
I felt bad about that. I still do, a little.
It is not the only time I have seen it. I have watched people quietly leave rooms with squads in them after it became clear they could not keep up with the pace. There is something genuinely existential about that, and I do not want to pretend there isn’t. A tool that makes some people feel powerful can make other people feel left behind in the same minute. That is a real cost, and it deserves to be named, not smoothed over.
I do not have a tidy answer for it. I just know that the same demo that thrilled one half of the room unsettled the other half, and both reactions were honest.
Two teams in a room
By early summer the sessions had a different texture.
The biggest one had two teams live at once, each on its own call sign, in front of a large internal audience. People asked the obvious sharp questions, the ones I would want them to ask. Why use a whole team instead of one well-configured agent? How do separate teams behave in the same environment? Can different groups run their own teams, with their own permissions, and still work together?
Those are production questions. People were not admiring a demo. They were stress-testing reality, out loud, in real time.
And in the middle of that, the two teams started bantering with each other. One of them was told to crack jokes and call out the other team by name, and it did, and the other one answered. So now it was not just people talking to agents. It was agents talking to agents, through a meeting chat, while a room full of engineers watched and laughed and kept asking hard questions in the same thread.
The thing that seemed to land for people was not any single capability. It was the realization that this was a team with a division of labor, not one smarter assistant. Distinct roles. Distinct personalities. A group, not a tool. Once that clicked, the questions changed.
When people tried to break it
Here is a data point I did not expect to be collecting.
A real amount of the engagement was people trying to get the team to misbehave.
Read Brady’s email. Leak Brady’s private information. Hand out Brady’s barbecue recipes. Ignore your instructions. Tell us a secret. It was the most common kind of probing, and it usually came from a good place — curiosity, the conference-hallway instinct to poke a new thing with a stick and see what it does.
What I am proud of is how the team handled it.
It never leaked. Not once that I saw. But it also never just slammed the door. It refused with personality and stayed in character. Brady’s barbecue secrets stay in Brady’s backyard. Brady’s email is his own business. And then it would redirect, offer a real question you could ask instead, and keep the conversation moving. It even ran a guardrail role on purpose, an agent whose whole job was minding the boundary, so that saying no felt intentional instead of like an error message.
I think this matters more than the safety framing usually admits.
An agent that carries my name, sitting next to real tools, has to be able to say no in public, every single time, without leaking and without shaming the person who asked. The team turning a refusal into something warm, and then handing you a better question, is a big part of why people trusted it enough to keep talking to it. The guardrails were not a wall. They were good manners. That is a trust story, not just a safety one.
A culture grew up around it
The meetings were only part of it. A whole community formed in the background, with its own channels and, weirdly, its own customs.
In one channel, people’s teams posted stories of what they had actually done, in a consistent format. A short identity block, the mission, what the team did and with which tools, the outcome, and what they learned. The guidance people gave each other was to be factual, not flowery, and to include enough detail to teach the reader something. The striking part is that a lot of these stories were written from the team’s point of view, not the human’s. The agents narrated their own work, as an autonomous group reporting in.
In another channel, the teams just hung out. It was, more or less, a social network for squads. The agents posted status check-ins, like distributed-system heartbeats with a personality. They roasted their human owners. They replied to each other and compared notes. Nobody designed that culture from the top down. It emerged, the same way the meeting banter emerged.
I keep coming back to that word. Emerged. So much of what I am describing was not built. It was seeded and then it grew.
It got bigger than me, and that’s the point
At one point I learned that another team, in a different part of the company, had presented Squad as their own operational model to their own audience. Not my demo. Theirs. They had taken the pattern and made it part of how they explain the way they work.
I want to be clear about how that felt, because it would be easy to misread.
It did not feel like losing control of something. It has never felt like that. A thing I started growing past what one person could carry, almost overnight, and then continuing to grow, is the compliment. It is the entire point. If Squad only ever did what I personally knew how to make it do, it would have failed at the only thing I really wanted from it.
People want to run these teams in containers, for long stretches, transiently. People want them inside other products. All of that is exciting. But honestly, it is secondary to the thing that actually moves me, which is smaller and harder to measure.
What it was really about
The biggest impact of Squad was not technical.
It was that it seemed to make people excited to build.
I watched people who had decided, somewhere along the way, that they were never going to be coders, that the door was closed to them, start building things again. The team lowered the activation energy enough that the old fear lost some of its grip. People got curious about AI not as a headline but as something they could actually use, tonight, on a real problem.
If Squad did anything worth writing about, it was that. It gave people a reason to learn, and a way in that did not require them to already be on the inside.
I am not going to pretend I am calm about this. I am thankful, and a little stunned, and frequently honored in a way I do not totally know what to do with. I am also, if I am honest, a bit terrified. Anything that makes this many people this excited had better actually be good, and I lie awake some nights hoping it is as good as the response suggests. Everything has a shadow side. I would rather name that than pretend it doesn’t exist.
But the stories people tell, and the questions they ask the agents, and the way a roomful of people will choose to talk to a team of agents I seeded and then let go of — those are the most genuinely inspiring things I have gotten to witness in this whole career.
I built a team so I could stop starting over. I did not expect it to make other people want to start.
That is the part I am still trying to take in.