Coding agents have changed how I work more than any tool I’ve adopted in the last decade. What they haven’t changed, however, is how isolated the work feels.
As a coder I’ve always valued two different modes. One is the zone: head down, one hard problem, hours focused on coding up the problem. The other is the room: five people, a whiteboard, and an idea that none of them walked in with. Today’s tools have made the first mode highly productive and satisfying. The second is roughly where it was a decade ago.
That gap isn’t small. Working out how people and agents operate side by side as a group is a design problem and a human factors problem as much as a technical one — and problems sitting in that overlap are the ones that intrigue me.
That curiosity has led me down a number of paths, and the useful thing about this moment is that exploring these ideas is cheap. Turning an idea into a prototype, and a prototype into something people can actually use, has never been faster or less expensive.
So this is where I’ll publish the results: a series of open-source experiments and the rationales behind them. Hopefully it nudges others into building on these ideas and into making the collaborative tools I’d like to use with my colleagues.
Ten geniuses, all agreeing with you
The first tool I’ve built is for brainstorming as a group — something that has quietly gotten harder, for a reason that took me a while to name.
Everyone now walks into the room with an agent that is an expert on exactly one person’s context: theirs. It has read their code, followed their reasoning, absorbed their assumptions. Picture a genius-level yes-man — brilliant, fully briefed, almost always in agreement with you.
Now picture ten of them in a room, each affirming a different person.
That is precisely the opposite of what a brainstorm is for. We get people together because no single person holds the whole picture — because the useful friction comes from someone carrying context you don’t have. Ten agents, each perfectly aligned with a different person, don’t produce ten perspectives in conversation. They produce ten confident monologues.
We used to do this with sticky notes, in rooms full of product owners, designers, business analysts and engineers, all providing their own perspective. The sticky note was the whole technology, and its most important property was that it was shared. The wall held the room’s thinking.
Today all of those perspectives are still there, and even more refined by our own agents. It’s all there in the room. It’s just siloed by default, and nothing we have today lets people break those silos open.
Riff
The idea I’ve been exploring is a simple take on this. Create an ephemeral session. Everyone in the room joins the session and can now contribute their ideas, refined by their agents, as they evolve. Everyone can share, refine and Riff on each other’s ideas, and use the shared session view to discuss them. Each person’s agent donates a context capsule to a shared board.
The capsule is the sticky note’s descendant. Where a sticky note held a jotting, a capsule holds structure:
| Field | What it holds |
|---|---|
objective | What this person is trying to solve |
approach | The angle they’re taking |
keyFindings | What their agent has surfaced so far |
openQuestions | What’s still unclear |
riffedFrom | The capsule this one was forked from |
A capsule isn’t just a static note, though. It carries lineage, and it carries the working context behind the idea — enough for another person’s agent to continue the thinking instead of reconstructing it from scratch.
The session board is the wall. Everyone’s objective, findings and open questions sit side by side, updating as their agent works. As each idea is refined, people can share it, see how it is evolving, and contribute their own knowledge back to it. The board depicts a live view of how ideas are evolving and coalescing as the team explores the problem space.
Riff is designed so that a person can seamlessly click on a context capsule and all of the relevant context gets automatically added to their own Claude Code session, waiting there for them to riff on. No copy-paste, no re-explaining, no “let me paste you my prompt.” The board records who riffed on whom, so the room can see how its thinking actually travelled.
The brainstorming loop becomes: brainstorm → share → discuss → riff → repeat.
It’s deliberately small to simplify the experience. At the end of a session, ideas can be synthesized, clustered, collected and refined in a shared session repository.
It runs in the room
I designed it to be secure and quick to start. Teams must be co-located on the same Wi-Fi to join a session: one person creates the session server, others join it. No accounts, no cloud, nothing leaves the building.
That isn’t a philosophical stance so much as a practical one. The board only works if people put real, unfinished work on it, and people are far more willing to do that when the thing holding it is a laptop they can see across the table.
Riff is an attempt to explore what collaborative brainstorming looks like when the agents can take part too — a temporary mind-meld that connects a room’s individual agent sessions and points them at a collective goal instead of ten private ones.
Trying it
You need Node 22 or newer. One person hosts the session:
npx riffboard startThat prints a URL, a join code, and a certificate fingerprint. Everyone else opens the URL on the same Wi-Fi, checks the fingerprint matches, and enters the code. To wire up your own Claude Code, click Connect Claude Code on the board and paste the single command it copies for you.
No accounts, no signup, no cloud — the whole thing is one process on somebody’s laptop.
Known limitations
- Brainstorming sessions are stored and summarized into a single markdown file.
- Local network support only — no remote participants.
- Claude Code is the only agent adapter supported so far.
What’s next
- Support for other agent harnesses and adapters.
- Remote network and participant support.
Riff is MIT licensed and available on GitHub. Contribute, extend it, fork it — riff on it, in the truest sense. Let’s see where it goes.