Pair programming is a rubber duck with a heartbeat. That is what it has always been for us: the duck listens, sometimes pushes back, sometimes just holds the space while you think. The thing nobody says out loud is that the duck also catches mistakes. Not every mistake — the kind that only surface when someone else has to look at what you just typed.
We have been building software together for a long time. The best sessions of our careers — the ones that stuck in memory, the ones that turned into the product we were most proud of — all looked the same: two chairs, one screen, one keyboard that gets passed back and forth, one shared whiteboard that nobody ever photographs. Those sessions used to happen in an office. For the last few years, they happened over video calls with a lot of "wait, can you scroll up a little?".
The itch
Somewhere around the point where Claude and GPT became real coding partners — not novelties, not demos, but tools we reached for dozens of times a day — the two sides of our work started to drift apart.
On one side: the live, synchronous, rubber-duck-with-a-heartbeat tradition. Two humans sharing a buffer, editing the same file, watching each other's cursor. On the other: the long, asynchronous, turn-taking dance with an agent in a side-panel chat. Paste some code, ask a question, read a reply, paste it somewhere else, repeat. Good answers, great sometimes, but the rhythm was wrong. The agent never sat in the chair.
We noticed it first on a Tuesday afternoon trying to shepherd a migration through. One of us was on the keyboard, one of us was watching, and Claude was open in a third window answering a question that was three steps behind. The agent could see everything we pasted but none of the state we were holding in our heads. It kept suggesting the thing we tried ten minutes ago and rejected. You could feel it in the room — we had three collaborators, and none of them had full presence.
Two dead ends
The obvious first thing is to make the agent smarter. More context window. Better retrieval. Smarter router. All real, all progress, but none of it fixed the asymmetry. The agent was still watching from the other side of a window.
The obvious second thing is to glue it into the editor. Copilot did this. Cursor did this. Every IDE extension does this. It is great. But it solves a different problem — the solo-coder-gets-autocomplete problem. When there are two humans in the room, the extension lives inside one editor. There is still an odd one out.
VS Code Live Share got closer. Two humans, same buffer, same cursor. But Live Share is a tunnel, not a session: your teammate sees your editor through a peephole, and the moment you disconnect the whole thing ends. There is no persistent plan, no shared terminal that outlives the window, no memory. And agents still lived in a panel, not in the session.
What we actually wanted
What we wanted was a room. A URL you could send someone. A live document that remembered what you were working on yesterday. A terminal that was the terminal, not a copy of someone else's. A chat that lived beside the code, not in a different browser tab. And — this was the thing that nobody else was building — an agent that could walk into the room and be a participant. Its own cursor. Its own avatar. Its own messages in the same chat. Playing by the same rules.
When we wrote down the list, it looked a lot like a design for a video game lobby. A persistent session that humans and agents both join. Everyone with presence. Everyone with roles. Everyone with a clear view of what anyone else had just done. That framing unlocked a lot. It is no accident that the product is called Huddle.
The bet
Our bet is that the future of the coding tool is not a better autocomplete. It is a better session. Agents are going to keep getting more capable, and the thing that separates the teams who ship with them from the teams who fight them is whether the agent is in the room or on the side.
We also bet that this only works if you own your keys. If Huddle made money on inference markup, we would have a reason to push you toward whatever model we made the fattest margin on. Instead, your Anthropic bill is your Anthropic bill, your OpenAI bill is your OpenAI bill, and we charge a flat seat price. Your agent is yours. Your bill is yours. Our incentive is to make the room you meet in feel like the best room you have ever worked in.
What is next
We are in beta. The product is useful today — two of us used it to write this post, with an agent drafting a second angle in a parallel session — and there is a long list of sharp edges still to file down. If that sounds interesting, grab a seat. If you want to read the shape of the thing first, the concepts page has the mental model and the getting-started guide has the first five minutes.
Either way — thanks for reading this far. Let us know what you think.